diff --git a/.gitignore b/.gitignore index 47e48237..43a89552 100644 --- a/.gitignore +++ b/.gitignore @@ -157,11 +157,20 @@ cython_debug/ # be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore # and can be added to the global gitignore or merged into this file. For a more nuclear # option (not recommended) you can uncomment the following to ignore the entire idea folder. -#.idea/ +.idea/ .vscode/ +# macos +*.DS_Store + # misc files +data/ dataset/ runs/ checkpoints/ outputs/ +samples/ +pretrained_models/ + +# Secret files +hostfile diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md new file mode 100644 index 00000000..b2ef579c --- /dev/null +++ b/CONTRIBUTING.md @@ -0,0 +1,91 @@ +# Contributing + +The Open-Sora project welcomes any constructive contribution from the community and the team is more than willing to work on problems you have encountered to make it a better project. + +## Development Environment Setup + +To contribute to Open-Sora, we would like to first guide you to set up a proper development environment so that you can better implement your code. You can install this library from source with the `editable` flag (`-e`, for development mode) so that your change to the source code will be reflected in runtime without re-installation. + +You can refer to the [Installation Section](./README.md#installation) and replace `pip install -v .` with `pip install -v -e .`. + + +### Code Style + +We have some static checks when you commit your code change, please make sure you can pass all the tests and make sure the coding style meets our requirements. We use pre-commit hook to make sure the code is aligned with the writing standard. To set up the code style checking, you need to follow the steps below. + +```shell +# these commands are executed under the Open-Sora directory +pip install pre-commit +pre-commit install +``` + +Code format checking will be automatically executed when you commit your changes. + + +## Contribution Guide + +You need to follow these steps below to make contribution to the main repository via pull request. You can learn about the details of pull request [here](https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/proposing-changes-to-your-work-with-pull-requests/about-pull-requests). + +### 1. Fork the Official Repository + +Firstly, you need to visit the [Open-Sora repository](https://github.com/hpcaitech/Open-Sora) and fork into your own account. The `fork` button is at the right top corner of the web page alongside with buttons such as `watch` and `star`. + +Now, you can clone your own forked repository into your local environment. + +```shell +git clone https://github.com//Open-Sora.git +``` + +### 2. Configure Git + +You need to set the official repository as your upstream so that you can synchronize with the latest update in the official repository. You can learn about upstream [here](https://www.atlassian.com/git/tutorials/git-forks-and-upstreams). + +Then add the original repository as upstream + +```shell +cd Open-Sora +git remote add upstream https://github.com/hpcaitech/Open-Sora.git +``` + +you can use the following command to verify that the remote is set. You should see both `origin` and `upstream` in the output. + +```shell +git remote -v +``` + +### 3. Synchronize with Official Repository + +Before you make changes to the codebase, it is always good to fetch the latest updates in the official repository. In order to do so, you can use the commands below. + +```shell +git fetch upstream +git checkout main +git merge upstream/main +git push origin main +``` + +### 5. Create a New Branch + +You should not make changes to the `main` branch of your forked repository as this might make upstream synchronization difficult. You can create a new branch with the appropriate name. General branch name format should start with `hotfix/` and `feature/`. `hotfix` is for bug fix and `feature` is for addition of a new feature. + + +```shell +git checkout -b +``` + +### 6. Implementation and Code Commit + +Now you can implement your code change in the source code. Remember that you installed the system in development, thus you do not need to uninstall and install to make the code take effect. The code change will be reflected in every new PyThon execution. +You can commit and push the changes to your local repository. The changes should be kept logical, modular and atomic. + +```shell +git add -A +git commit -m "" +git push -u origin +``` + +### 7. Open a Pull Request + +You can now create a pull request on the GitHub webpage of your repository. The source branch is `` of your repository and the target branch should be `main` of `hpcaitech/Open-Sora`. After creating this pull request, you should be able to see it [here](https://github.com/hpcaitech/Open-Sora/pulls). + +The Open-Sora team will review your code change and merge your code if applicable. diff --git a/LICENSE b/LICENSE index 261eeb9e..7327c123 100644 --- a/LICENSE +++ b/LICENSE @@ -199,3 +199,483 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. + + ========================================================================= + This project is inspired by the listed projects and is subject to the following licenses: + + 1. Latte (https://github.com/Vchitect/Latte/blob/main/LICENSE) + + Copyright 2024 Latte + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. + + 2. PixArt-alpha (https://github.com/PixArt-alpha/PixArt-alpha/blob/master/LICENSE) + + Copyright (C) 2024 PixArt-alpha/PixArt-alpha + + This program is free software: you can redistribute it and/or modify + it under the terms of the GNU Affero General Public License as published + by the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU Affero General Public License for more details. + + You should have received a copy of the GNU Affero General Public License + along with this program. If not, see . + + 3. dpm-solver (https://github.com/LuChengTHU/dpm-solver/blob/main/LICENSE) + + MIT License + + Copyright (c) 2022 Cheng Lu + + Permission is hereby granted, free of charge, to any person obtaining a copy + of this software and associated documentation files (the "Software"), to deal + in the Software without restriction, including without limitation the rights + to use, copy, modify, merge, publish, distribute, sublicense, and/or sell + copies of the Software, and to permit persons to whom the Software is + furnished to do so, subject to the following conditions: + + The above copyright notice and this permission notice shall be included in all + copies or substantial portions of the Software. + + THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + SOFTWARE. + + 4. DiT (https://github.com/facebookresearch/DiT/blob/main/LICENSE.txt) + + Attribution-NonCommercial 4.0 International + + ======================================================================= + + Creative Commons Corporation ("Creative Commons") is not a law firm and + does not provide legal services or legal advice. Distribution of + Creative Commons public licenses does not create a lawyer-client or + other relationship. Creative Commons makes its licenses and related + information available on an "as-is" basis. Creative Commons gives no + warranties regarding its licenses, any material licensed under their + terms and conditions, or any related information. Creative Commons + disclaims all liability for damages resulting from their use to the + fullest extent possible. + + Using Creative Commons Public Licenses + + Creative Commons public licenses provide a standard set of terms and + conditions that creators and other rights holders may use to share + original works of authorship and other material subject to copyright + and certain other rights specified in the public license below. The + following considerations are for informational purposes only, are not + exhaustive, and do not form part of our licenses. + + Considerations for licensors: Our public licenses are + intended for use by those authorized to give the public + permission to use material in ways otherwise restricted by + copyright and certain other rights. Our licenses are + irrevocable. Licensors should read and understand the terms + and conditions of the license they choose before applying it. + Licensors should also secure all rights necessary before + applying our licenses so that the public can reuse the + material as expected. Licensors should clearly mark any + material not subject to the license. This includes other CC- + licensed material, or material used under an exception or + limitation to copyright. More considerations for licensors: + wiki.creativecommons.org/Considerations_for_licensors + + Considerations for the public: By using one of our public + licenses, a licensor grants the public permission to use the + licensed material under specified terms and conditions. If + the licensor's permission is not necessary for any reason--for + example, because of any applicable exception or limitation to + copyright--then that use is not regulated by the license. Our + licenses grant only permissions under copyright and certain + other rights that a licensor has authority to grant. Use of + the licensed material may still be restricted for other + reasons, including because others have copyright or other + rights in the material. A licensor may make special requests, + such as asking that all changes be marked or described. + Although not required by our licenses, you are encouraged to + respect those requests where reasonable. More_considerations + for the public: + wiki.creativecommons.org/Considerations_for_licensees + + ======================================================================= + + Creative Commons Attribution-NonCommercial 4.0 International Public + License + + By exercising the Licensed Rights (defined below), You accept and agree + to be bound by the terms and conditions of this Creative Commons + Attribution-NonCommercial 4.0 International Public License ("Public + License"). To the extent this Public License may be interpreted as a + contract, You are granted the Licensed Rights in consideration of Your + acceptance of these terms and conditions, and the Licensor grants You + such rights in consideration of benefits the Licensor receives from + making the Licensed Material available under these terms and + conditions. + + Section 1 -- Definitions. + + a. Adapted Material means material subject to Copyright and Similar + Rights that is derived from or based upon the Licensed Material + and in which the Licensed Material is translated, altered, + arranged, transformed, or otherwise modified in a manner requiring + permission under the Copyright and Similar Rights held by the + Licensor. For purposes of this Public License, where the Licensed + Material is a musical work, performance, or sound recording, + Adapted Material is always produced where the Licensed Material is + synched in timed relation with a moving image. + + b. Adapter's License means the license You apply to Your Copyright + and Similar Rights in Your contributions to Adapted Material in + accordance with the terms and conditions of this Public License. + + c. Copyright and Similar Rights means copyright and/or similar rights + closely related to copyright including, without limitation, + performance, broadcast, sound recording, and Sui Generis Database + Rights, without regard to how the rights are labeled or + categorized. For purposes of this Public License, the rights + specified in Section 2(b)(1)-(2) are not Copyright and Similar + Rights. + d. Effective Technological Measures means those measures that, in the + absence of proper authority, may not be circumvented under laws + fulfilling obligations under Article 11 of the WIPO Copyright + Treaty adopted on December 20, 1996, and/or similar international + agreements. + + e. Exceptions and Limitations means fair use, fair dealing, and/or + any other exception or limitation to Copyright and Similar Rights + that applies to Your use of the Licensed Material. + + f. Licensed Material means the artistic or literary work, database, + or other material to which the Licensor applied this Public + License. + + g. Licensed Rights means the rights granted to You subject to the + terms and conditions of this Public License, which are limited to + all Copyright and Similar Rights that apply to Your use of the + Licensed Material and that the Licensor has authority to license. + + h. Licensor means the individual(s) or entity(ies) granting rights + under this Public License. + + i. NonCommercial means not primarily intended for or directed towards + commercial advantage or monetary compensation. For purposes of + this Public License, the exchange of the Licensed Material for + other material subject to Copyright and Similar Rights by digital + file-sharing or similar means is NonCommercial provided there is + no payment of monetary compensation in connection with the + exchange. + + j. Share means to provide material to the public by any means or + process that requires permission under the Licensed Rights, such + as reproduction, public display, public performance, distribution, + dissemination, communication, or importation, and to make material + available to the public including in ways that members of the + public may access the material from a place and at a time + individually chosen by them. + + k. Sui Generis Database Rights means rights other than copyright + resulting from Directive 96/9/EC of the European Parliament and of + the Council of 11 March 1996 on the legal protection of databases, + as amended and/or succeeded, as well as other essentially + equivalent rights anywhere in the world. + + l. You means the individual or entity exercising the Licensed Rights + under this Public License. Your has a corresponding meaning. + + Section 2 -- Scope. + + a. License grant. + + 1. Subject to the terms and conditions of this Public License, + the Licensor hereby grants You a worldwide, royalty-free, + non-sublicensable, non-exclusive, irrevocable license to + exercise the Licensed Rights in the Licensed Material to: + + a. reproduce and Share the Licensed Material, in whole or + in part, for NonCommercial purposes only; and + + b. produce, reproduce, and Share Adapted Material for + NonCommercial purposes only. + + 2. Exceptions and Limitations. For the avoidance of doubt, where + Exceptions and Limitations apply to Your use, this Public + License does not apply, and You do not need to comply with + its terms and conditions. + + 3. Term. The term of this Public License is specified in Section + 6(a). + + 4. Media and formats; technical modifications allowed. The + Licensor authorizes You to exercise the Licensed Rights in + all media and formats whether now known or hereafter created, + and to make technical modifications necessary to do so. The + Licensor waives and/or agrees not to assert any right or + authority to forbid You from making technical modifications + necessary to exercise the Licensed Rights, including + technical modifications necessary to circumvent Effective + Technological Measures. For purposes of this Public License, + simply making modifications authorized by this Section 2(a) + (4) never produces Adapted Material. + + 5. Downstream recipients. + + a. Offer from the Licensor -- Licensed Material. Every + recipient of the Licensed Material automatically + receives an offer from the Licensor to exercise the + Licensed Rights under the terms and conditions of this + Public License. + + b. No downstream restrictions. You may not offer or impose + any additional or different terms or conditions on, or + apply any Effective Technological Measures to, the + Licensed Material if doing so restricts exercise of the + Licensed Rights by any recipient of the Licensed + Material. + + 6. No endorsement. Nothing in this Public License constitutes or + may be construed as permission to assert or imply that You + are, or that Your use of the Licensed Material is, connected + with, or sponsored, endorsed, or granted official status by, + the Licensor or others designated to receive attribution as + provided in Section 3(a)(1)(A)(i). + + b. Other rights. + + 1. Moral rights, such as the right of integrity, are not + licensed under this Public License, nor are publicity, + privacy, and/or other similar personality rights; however, to + the extent possible, the Licensor waives and/or agrees not to + assert any such rights held by the Licensor to the limited + extent necessary to allow You to exercise the Licensed + Rights, but not otherwise. + + 2. Patent and trademark rights are not licensed under this + Public License. + + 3. To the extent possible, the Licensor waives any right to + collect royalties from You for the exercise of the Licensed + Rights, whether directly or through a collecting society + under any voluntary or waivable statutory or compulsory + licensing scheme. In all other cases the Licensor expressly + reserves any right to collect such royalties, including when + the Licensed Material is used other than for NonCommercial + purposes. + + Section 3 -- License Conditions. + + Your exercise of the Licensed Rights is expressly made subject to the + following conditions. + + a. Attribution. + + 1. If You Share the Licensed Material (including in modified + form), You must: + + a. retain the following if it is supplied by the Licensor + with the Licensed Material: + + i. identification of the creator(s) of the Licensed + Material and any others designated to receive + attribution, in any reasonable manner requested by + the Licensor (including by pseudonym if + designated); + + ii. a copyright notice; + + iii. a notice that refers to this Public License; + + iv. a notice that refers to the disclaimer of + warranties; + + v. a URI or hyperlink to the Licensed Material to the + extent reasonably practicable; + + b. indicate if You modified the Licensed Material and + retain an indication of any previous modifications; and + + c. indicate the Licensed Material is licensed under this + Public License, and include the text of, or the URI or + hyperlink to, this Public License. + + 2. You may satisfy the conditions in Section 3(a)(1) in any + reasonable manner based on the medium, means, and context in + which You Share the Licensed Material. For example, it may be + reasonable to satisfy the conditions by providing a URI or + hyperlink to a resource that includes the required + information. + + 3. If requested by the Licensor, You must remove any of the + information required by Section 3(a)(1)(A) to the extent + reasonably practicable. + + 4. If You Share Adapted Material You produce, the Adapter's + License You apply must not prevent recipients of the Adapted + Material from complying with this Public License. + + Section 4 -- Sui Generis Database Rights. + + Where the Licensed Rights include Sui Generis Database Rights that + apply to Your use of the Licensed Material: + + a. for the avoidance of doubt, Section 2(a)(1) grants You the right + to extract, reuse, reproduce, and Share all or a substantial + portion of the contents of the database for NonCommercial purposes + only; + + b. if You include all or a substantial portion of the database + contents in a database in which You have Sui Generis Database + Rights, then the database in which You have Sui Generis Database + Rights (but not its individual contents) is Adapted Material; and + + c. You must comply with the conditions in Section 3(a) if You Share + all or a substantial portion of the contents of the database. + + For the avoidance of doubt, this Section 4 supplements and does not + replace Your obligations under this Public License where the Licensed + Rights include other Copyright and Similar Rights. + + Section 5 -- Disclaimer of Warranties and Limitation of Liability. + + a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE + EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS + AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF + ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS, + IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION, + WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR + PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS, + ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT + KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT + ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU. + + b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE + TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION, + NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT, + INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES, + COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR + USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN + ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR + DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR + IN PART, THIS LIMITATION MAY NOT APPLY TO YOU. + + c. The disclaimer of warranties and limitation of liability provided + above shall be interpreted in a manner that, to the extent + possible, most closely approximates an absolute disclaimer and + waiver of all liability. + + Section 6 -- Term and Termination. + + a. This Public License applies for the term of the Copyright and + Similar Rights licensed here. However, if You fail to comply with + this Public License, then Your rights under this Public License + terminate automatically. + + b. Where Your right to use the Licensed Material has terminated under + Section 6(a), it reinstates: + + 1. automatically as of the date the violation is cured, provided + it is cured within 30 days of Your discovery of the + violation; or + + 2. upon express reinstatement by the Licensor. + + For the avoidance of doubt, this Section 6(b) does not affect any + right the Licensor may have to seek remedies for Your violations + of this Public License. + + c. For the avoidance of doubt, the Licensor may also offer the + Licensed Material under separate terms or conditions or stop + distributing the Licensed Material at any time; however, doing so + will not terminate this Public License. + + d. Sections 1, 5, 6, 7, and 8 survive termination of this Public + License. + + Section 7 -- Other Terms and Conditions. + + a. The Licensor shall not be bound by any additional or different + terms or conditions communicated by You unless expressly agreed. + + b. Any arrangements, understandings, or agreements regarding the + Licensed Material not stated herein are separate from and + independent of the terms and conditions of this Public License. + + Section 8 -- Interpretation. + + a. For the avoidance of doubt, this Public License does not, and + shall not be interpreted to, reduce, limit, restrict, or impose + conditions on any use of the Licensed Material that could lawfully + be made without permission under this Public License. + + b. To the extent possible, if any provision of this Public License is + deemed unenforceable, it shall be automatically reformed to the + minimum extent necessary to make it enforceable. If the provision + cannot be reformed, it shall be severed from this Public License + without affecting the enforceability of the remaining terms and + conditions. + + c. No term or condition of this Public License will be waived and no + failure to comply consented to unless expressly agreed to by the + Licensor. + + d. Nothing in this Public License constitutes or may be interpreted + as a limitation upon, or waiver of, any privileges and immunities + that apply to the Licensor or You, including from the legal + processes of any jurisdiction or authority. + + ======================================================================= + + Creative Commons is not a party to its public + licenses. Notwithstanding, Creative Commons may elect to apply one of + its public licenses to material it publishes and in those instances + will be considered the “Licensor.” The text of the Creative Commons + public licenses is dedicated to the public domain under the CC0 Public + Domain Dedication. Except for the limited purpose of indicating that + material is shared under a Creative Commons public license or as + otherwise permitted by the Creative Commons policies published at + creativecommons.org/policies, Creative Commons does not authorize the + use of the trademark "Creative Commons" or any other trademark or logo + of Creative Commons without its prior written consent including, + without limitation, in connection with any unauthorized modifications + to any of its public licenses or any other arrangements, + understandings, or agreements concerning use of licensed material. For + the avoidance of doubt, this paragraph does not form part of the + public licenses. + + Creative Commons may be contacted at creativecommons.org. + + 5. OpenDiT (https://github.com/NUS-HPC-AI-Lab/OpenDiT/blob/master/LICENSE) + + Copyright OpenDiT + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. + + diff --git a/README.md b/README.md index 46d75c23..4a1839c4 100644 --- a/README.md +++ b/README.md @@ -1,165 +1,207 @@ -# 🎥 Open-Sora -
- - [![GitHub Repo stars](https://img.shields.io/github/stars/hpcaitech/Open-Sora?style=social)](https://github.com/hpcaitech/Open-Sora/stargazers) - [![slack badge](https://img.shields.io/badge/Slack-join-blueviolet?logo=slack&)](https://github.com/hpcaitech/public_assets/tree/main/colossalai/contact/slack) - [![WeChat badge](https://img.shields.io/badge/微信-加入-green?logo=wechat&)](https://raw.githubusercontent.com/hpcaitech/public_assets/main/colossalai/img/WeChat.png) - -
+

+ +

-## 📎 Table of Contents - -- [🎥 Open-Sora](#-open-sora) - - [📎 Table of Contents](#-table-of-contents) - - [📍 Overview](#-overview) - - [📂 Dataset Preparation](#-dataset-preparation) - - [Use MSR-VTT](#use-msr-vtt) - - [Use Customized Datasets](#use-customized-datasets) - - [🚀 Get Started](#-get-started) - - [Training](#training) - - [Inference](#inference) - - [🪄 Acknowledgement](#-acknowledgement) - -## Latest News -* [2024/03] [Open-Sora：Sora Replication Solution with 46% Cost Reduction, Sequence Expansion to Nearly a Million](https://hpc-ai.com/blog/open-sora) - -## 📍 Overview -Open-Sora is an open-source project that provides a high-performance implementation of the development pipeline that Sora might use powered by [Colossal-AI](https://github.com/hpcaitech/ColossalAI), including: - -- Provides **a complete Sora reproduction architecture solution**, including the whole process from data processing to training and deployment. -- Supports **dynamic resolution**, training can directly train any resolution of the video, without scaling. -- Supports **multiple model structures**. Since the actual model structure of Sora is unknown, we implement three common multimodal model structures such as adaLN-zero, cross attention, and in-context conditioning (token concat). -- Supports **multiple video compression methods**. Users can choose to use original video, VQVAE (video native model), or SD-VAE (image native model) for training. -- Supports **multiple parallel training optimizations**. Including the AI large model system optimization capability compatible with Colossal-AI, and hybrid sequence parallelism with Ulysses and FastSeq. - -

-

+
+ + + +
-

- -

+## Open-Sora: Democratizing Efficient Video Production for All +We present **Open-Sora**, an initiative dedicated to **efficiently** produce high-quality video and make the model, +tools and contents accessible to all. By embracing **open-source** principles, +Open-Sora not only democratizes access to advanced video generation techniques, but also offers a +streamlined and user-friendly platform that simplifies the complexities of video production. +With Open-Sora, we aim to inspire innovation, creativity, and inclusivity in the realm of content creation. [[中文]](/docs/README_zh.md) -

- -

+## 📰 News -## 📂 Dataset Preparation +* **[2024.03.18]** 🔥 We release **Open-Sora 1.0**, an fully open-sourced project for video generation. +Open-Sora 1.0 supports a full pipeline of video data preprocessing, training with + acceleration, +inference, and more. Our provided [checkpoints](#model-weights) can produce 2~5s 512x512 videos with only 3 days training. +* **[2024.03.04]** Open-Sora provides training with 46% cost reduction. -### Use MSR-VTT +## 🎥 Latest Demo -We use [MSR-VTT](https://cove.thecvf.com/datasets/839) dataset, which is a large-scale video description dataset. Users should preprocess the raw videos before training the model. You can use the following scripts to perform data processing. +| **2s 512×512** | **2s 512×512** | **2s 512×512** | +| ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------- | +| [](https://github.com/hpcaitech/Open-Sora/assets/99191637/de1963d3-b43b-4e68-a670-bb821ebb6f80) | [](https://github.com/hpcaitech/Open-Sora/assets/99191637/13f8338f-3d42-4b71-8142-d234fbd746cc) | [](https://github.com/hpcaitech/Open-Sora/assets/99191637/fa6a65a6-e32a-4d64-9a9e-eabb0ebb8c16) | +| A serene night scene in a forested area. [...] The video is a time-lapse, capturing the transition from day to night, with the lake and forest serving as a constant backdrop. | A soaring drone footage captures the majestic beauty of a coastal cliff, [...] The water gently laps at the rock base and the greenery that clings to the top of the cliff. | The majestic beauty of a waterfall cascading down a cliff into a serene lake. [...] The camera angle provides a bird's eye view of the waterfall. | +| | | | +| A bustling city street at night, filled with the glow of car headlights and the ambient light of streetlights. [...] | The vibrant beauty of a sunflower field. The sunflowers are arranged in neat rows, creating a sense of order and symmetry. [...] | A serene underwater scene featuring a sea turtle swimming through a coral reef. The turtle, with its greenish-brown shell [...] | +Videos are downsampled to `.gif` for display. Click for original videos. Texts are trimmed for display, see [here](/assets/texts/t2v_samples.txt) for full texts. See more samples at our [gallery](https://hpcaitech.github.io/Open-Sora/). -```bash -# Step 1: download the dataset to ./dataset/MSRVTT -bash scripts/data/download_msr_vtt_dataset.sh +## 🔆 New Features/Updates -# Step 2: collate the video and annotations -python scripts/data/collate_msr_vtt_dataset.py -d ./dataset/MSRVTT/ -o ./dataset/MSRVTT-collated +* 📍 Open-Sora-v1 released. Model weights are available [here](#model-weights). With only 400K video clips and 200 H800 days (compared with 152M samples in Stable Video Diffusion), we are able to generate 2s 512×512 videos. +* ✅ Three stages training from an image diffusion model to a video diffusion model. We provide the weights for each stage. +* ✅ Support training acceleration including accelerated transformer, faster T5 and VAE, and sequence parallelism. Open-Sora improve **55%** training speed when training on 64x512x512 videos. Details locates at [acceleration.md](docs/acceleration.md). +* ✅ We provide video cutting and captioning tools for data preprocessing. Instructions can be found [here](tools/data/README.md) and our data collection plan can be found at [datasets.md](docs/datasets.md). +* ✅ We find VQ-VAE from [VideoGPT](https://wilson1yan.github.io/videogpt/index.html) has a low quality and thus adopt a better VAE from [Stability-AI](https://huggingface.co/stabilityai/sd-vae-ft-mse-original). We also find patching in the time dimension deteriorates the quality. See our **[report](docs/report_v1.md)** for more discussions. +* ✅ We investigate different architectures including DiT, Latte, and our proposed STDiT. Our **STDiT** achieves a better trade-off between quality and speed. See our **[report](docs/report_v1.md)** for more discussions. +* ✅ Support clip and T5 text conditioning. +* ✅ By viewing images as one-frame videos, our project supports training DiT on both images and videos (e.g., ImageNet & UCF101). See [command.md](docs/command.md) for more instructions. +* ✅ Support inference with official weights from [DiT](https://github.com/facebookresearch/DiT), [Latte](https://github.com/Vchitect/Latte), and [PixArt](https://pixart-alpha.github.io/). -# Step 3: perform data processing -# NOTE: each script could several minutes so we apply the script to the dataset split individually -python scripts/data/preprocess_data.py -c ./dataset/MSRVTT-collated/train/annotations.json -v ./dataset/MSRVTT-collated/train/videos -o ./dataset/MSRVTT-processed/train -python scripts/data/preprocess_data.py -c ./dataset/MSRVTT-collated/val/annotations.json -v ./dataset/MSRVTT-collated/val/videos -o ./dataset/MSRVTT-processed/val -python scripts/data/preprocess_data.py -c ./dataset/MSRVTT-collated/test/annotations.json -v ./dataset/MSRVTT-collated/test/videos -o ./dataset/MSRVTT-processed/test -``` +
+View more -**If you want to use adaLN-zero, you should use `--use_pooled_text` when running `preprocess_data.py`** +* ✅ Refactor the codebase. See [structure.md](docs/structure.md) to learn the project structure and how to use the config files. -```bash +
-After completing these steps, you should have a processed MSR-VTT dataset in `./dataset/MSRVTT-processed`. +### TODO list sorted by priority +* [ ] Complete the data processing pipeline (including dense optical flow, aesthetics scores, text-image similarity, deduplication, etc.). See [datasets.md](/docs/datasets.md) for more information. **[WIP]** +* [ ] Training Video-VAE. **[WIP]** -### Use Customized Datasets +
+View more -You can also use other datasets and transform the dataset to the required format. You should prepare a captions file and a video directory. The captions file should be a JSON file or a JSONL file. The video directory should contain all the videos. +* [ ] Support image and video conditioning. +* [ ] Evaluation pipeline. +* [ ] Incoporate a better scheduler, e.g., rectified flow in SD3. +* [ ] Support variable aspect ratios, resolutions, durations. +* [ ] Support SD3 when released. -Here is an example of the captions file: +
-```json -[ - { - "file": "video0.mp4", - "captions": ["a girl is throwing away folded clothes", "a girl throwing cloths around"] - }, - { - "file": "video1.mp4", - "captions": ["a comparison of two opposing team football athletes"] - } -] -``` +## Contentss -Here is an example of the video directory: +* [Installation](#installation) +* [Model Weights](#model-weights) +* [Inference](#inference) +* [Data Processing](#data-processing) +* [Training](#training) +* [Contribution](#contribution) +* [Acknowledgement](#acknowledgement) +* [Citation](#citation) -``` -. -├── video0.mp4 -├── video1.mp4 -└── ... -``` +## Installation -Each video may have multiple captions. So the outputs are video-caption pairs. E.g., If the first video has two captions, then the output will be two video-caption pairs. +```bash +# create a virtual env +conda create -n opensora python=3.10 -We use [VQ-VAE](https://github.com/wilson1yan/VideoGPT/) to quantize the video frames. And we use [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#clip) to extract the text features. +# install torch +# the command below is for CUDA 12.1, choose install commands from +# https://pytorch.org/get-started/locally/ based on your own CUDA version +pip3 install torch torchvision -The output is an arrow dataset, which contains the following columns: "video_file", "video_latent_states", "text_latent_states". The dimension of "video_latent_states" is (T, H, W), and the dimension of "text_latent_states" is (S, D). +# install flash attention (optional) +pip install packaging ninja +pip install flash-attn --no-build-isolation -Then you can run the data processing script with the command below: +# install apex (optional) +pip install -v --disable-pip-version-check --no-cache-dir --no-build-isolation --config-settings "--build-option=--cpp_ext" --config-settings "--build-option=--cuda_ext" git+https://github.com/NVIDIA/apex.git -```bash -python preprocess_data.py -c /path/to/captions.json -v /path/to/video_dir -o /path/to/output_dir +# install xformers +pip3 install -U xformers --index-url https://download.pytorch.org/whl/cu121 + +# install this project +git clone https://github.com/hpcaitech/Open-Sora +cd Open-Sora +pip install -v . ``` -**If you want to use adaLN-zero, you should use `--use_pooled_text` when running `preprocess_data.py`** +After installation, we suggest reading [structure.md](docs/structure.md) to learn the project structure and how to use the config files. + +## Model Weights -Note that this script needs to be run on a machine with a GPU. To avoid CUDA OOM, we filter out the videos that are too long. +| Resoluion | Data | #iterations | Batch Size | GPU days (H800) | URL | +| ---------- | ------ | ----------- | ---------- | --------------- | ---------- | +| 16×256×256 | 366K | 80k | 8×64 | 117 | [:link:]() | +| 16×256×256 | 20K HQ | 24k | 8×64 | 45 | [:link:]() | +| 16×512×512 | 20K HQ | 20k | 2×64 | 35 | [:link:]() | +| 64×512×512 | 50K HQ | | | | TBD | +Our model's weight is partially initialized from [PixArt-α](https://github.com/PixArt-alpha/PixArt-alpha). The number of parameters is 724M. More information about training can be found in our **[report](/docs/report_v1.md)**. More about dataset can be found in [dataset.md](/docs/dataset.md). HQ means high quality. -## 🚀 Get Started +:warning: **LIMITATION**: Our model is trained on a limited budget. The quality and text alignment is relatively poor. The model performs badly especially on generating human beings and cannot follow detailed instructions. We are working on improving the quality and text alignment. -In this section, we will guide how to run training and inference. Before that, make sure you installed the dependencies with the command below. +## Inference + +To run inference with our provided weights, first download [T5](https://huggingface.co/DeepFloyd/t5-v1_1-xxl/tree/main) weights into `pretrained_models/t5_ckpts/t5-v1_1-xxl`. Then download the model weights. Run the following commands to generate samples. See [here](docs/structure.md#inference-config-demos) to customize the configuration. ```bash -pip install -r requirements.txt -``` +# Sample 16x256x256 (5s/sample) +torchrun --standalone --nproc_per_node 1 scripts/inference.py configs/opensora/inference/16x256x256.py --ckpt-path ./path/to/your/ckpt.pth -### Training +# Sample 16x512x512 (20s/sample, 100 time steps) +torchrun --standalone --nproc_per_node 1 scripts/inference.py configs/opensora/inference/16x512x512.py --ckpt-path ./path/to/your/ckpt.pth -You can invoke the training via the command below. +# Sample 64x512x512 (40s/sample, 100 time steps) +torchrun --standalone --nproc_per_node 1 scripts/inference.py configs/opensora/inference/64x512x512.py --ckpt-path ./path/to/your/ckpt.pth -```bash -bash ./scripts/train.sh +# Sample 64x512x512 with sequence parallelism (30s/sample, 100 time steps) +# sequence parallelism is enabled automatically when nproc_per_node is larger than 1 +torchrun --standalone --nproc_per_node 2 scripts/inference.py configs/opensora/inference/64x512x512.py --ckpt-path ./path/to/your/ckpt.pth ``` -You can also modify the arguments in `train.sh` for your own need. +The speed is tested on H800 GPUs. For inference with other models, see [here](docs/commands.md) for more instructions. + +## Data Processing -We provide three model architectures: "adaln", "cross-attn" and "token-concat". You can specify the model architecture with the `-x` option. +High-quality Data is the key to high-quality models. Our used datasets and data collection plan is [here](/docs/datasets.md). We provide tools to process video data. Currently, our data processing pipeline includes the following steps: -### Inference +1. Downloading datasets. [[docs](/tools/datasets/README.md)] +2. Split videos into clips. [[docs](/tools/scenedetect/README.md)] +3. Generate video captions. [[docs](/tools/caption/README.md)] -We've provided a script to perform inference, allowing you to generate videos from the trained model. You can invoke the inference via the command below. +## Training + +To launch training, first download [T5](https://huggingface.co/DeepFloyd/t5-v1_1-xxl/tree/main) weights into `pretrained_models/t5_ckpts/t5-v1_1-xxl`. Then run the following commands to launch training on a single node. ```bash -python sample.py -m "DiT/XL-2" --text "a person is walking on the street" --ckpt /path/to/checkpoint --height 256 --width 256 --fps 10 --sec 5 --disable-cfg +# 1 GPU, 16x256x256 +torchrun --nnodes=1 --nproc_per_node=1 scripts/train.py configs/opensora/train/16x256x512.py --data-path YOUR_CSV_PATH +# 8 GPUs, 64x512x512 +torchrun --nnodes=1 --nproc_per_node=8 scripts/train.py configs/opensora/train/64x512x512.py --data-path YOUR_CSV_PATH --ckpt-path YOUR_PRETRAINED_CKPT ``` -This will generate a "sample.mp4" file in the current directory. -For more command line options, you can use the following command to check the help message. +To launch training on multiple nodes, prepare a hostfile according to [ColossalAI](https://colossalai.org/docs/basics/launch_colossalai/#launch-with-colossal-ai-cli), and run the following commands. ```bash -python sample.py -h +colossalai run --nproc_per_node 8 --hostfile hostfile scripts/train.py configs/opensora/train/64x512x512.py --data-path YOUR_CSV_PATH --ckpt-path YOUR_PRETRAINED_CKPT ``` -Please make sure the video compressor and model architecture are consistent with the training settings. +For training other models and advanced usage, see [here](docs/commands.md) for more instructions. + +## Contribution + +If you wish to contribute to this project, you can refer to the [Contribution Guideline](./CONTRIBUTING.md). + +## Acknowledgement + +* [DiT](https://github.com/facebookresearch/DiT): Scalable Diffusion Models with Transformers. +* [OpenDiT](https://github.com/NUS-HPC-AI-Lab/OpenDiT): An acceleration for DiT training. We adopt valuable acceleration strategies for training progress from OpenDiT. +* [PixArt](https://github.com/PixArt-alpha/PixArt-alpha): An open-source DiT-based text-to-image model. +* [Latte](https://github.com/Vchitect/Latte): An attempt to efficiently train DiT for video. +* [StabilityAI VAE](https://huggingface.co/stabilityai/sd-vae-ft-mse-original): A powerful image VAE model. +* [CLIP](https://github.com/openai/CLIP): A powerful text-image embedding model. +* [T5](https://github.com/google-research/text-to-text-transfer-transformer): A powerful text encoder. +* [LLaVA](https://github.com/haotian-liu/LLaVA): A powerful image captioning model based on [Yi-34B](https://huggingface.co/01-ai/Yi-34B). + +We are grateful for their exceptional work and generous contribution to open source. + +## Citation + +```bibtex +@software{opensora, + author = {Zangwei Zheng and Xiangyu Peng and Yang You}, + title = {Open-Sora: Democratizing Efficient Video Production for All}, + month = {March}, + year = {2024}, + url = {https://github.com/hpcaitech/Open-Sora} +} +``` -## 🪄 Acknowledgement +[Zangwei Zheng](https://github.com/zhengzangw) and [Xiangyu Peng](https://github.com/xyupeng) equally contributed to this work during their internship at [HPC-AI Tech](https://hpc-ai.com/). -During the development of the project, we learned a lot from the following information: +## Star History -- [OpenAI Sora Technical Report](https://openai.com/research/video-generation-models-as-world-simulators) -- [VideoGPT Project](https://github.com/wilson1yan/VideoGPT) -- [Diffusion Transformers](https://github.com/facebookresearch/DiT) -- [Deepspeed Ulysses](https://arxiv.org/abs/2309.14509) -- [OpenDiT](https://github.com/NUS-HPC-AI-Lab/OpenDiT) +[![Star History Chart](https://api.star-history.com/svg?repos=hpcaitech/Open-Sora&type=Date)](https://star-history.com/#hpcaitech/Open-Sora&Date) diff --git a/assets/images/imagenet/train/n01440764/n01440764_10026.JPEG b/assets/images/imagenet/train/n01440764/n01440764_10026.JPEG new file mode 100644 index 00000000..b985769e Binary files /dev/null and b/assets/images/imagenet/train/n01440764/n01440764_10026.JPEG differ diff --git a/assets/images/imagenet/val/n01440764/ILSVRC2012_val_00000293.JPEG b/assets/images/imagenet/val/n01440764/ILSVRC2012_val_00000293.JPEG new file mode 100644 index 00000000..1b332471 Binary files /dev/null and b/assets/images/imagenet/val/n01440764/ILSVRC2012_val_00000293.JPEG differ diff --git a/assets/readme/colossal_ai.png b/assets/readme/colossal_ai.png new file mode 100644 index 00000000..05df34a6 Binary files /dev/null and b/assets/readme/colossal_ai.png differ diff --git a/assets/readme/icon_new.png b/assets/readme/icon_new.png new file mode 100644 index 00000000..096212ef Binary files /dev/null and b/assets/readme/icon_new.png differ diff --git a/assets/readme/icon_old.png b/assets/readme/icon_old.png new file mode 100644 index 00000000..7cae09bf Binary files /dev/null and b/assets/readme/icon_old.png differ diff --git a/assets/readme/sample_0.gif b/assets/readme/sample_0.gif new file mode 100644 index 00000000..a8fd9cfb Binary files /dev/null and b/assets/readme/sample_0.gif differ diff --git a/assets/readme/sample_1.gif b/assets/readme/sample_1.gif new file mode 100644 index 00000000..a4bc6394 Binary files /dev/null and b/assets/readme/sample_1.gif differ diff --git a/assets/readme/sample_2.gif b/assets/readme/sample_2.gif new file mode 100644 index 00000000..cea86acb Binary files /dev/null and b/assets/readme/sample_2.gif differ diff --git a/assets/readme/sample_3.gif b/assets/readme/sample_3.gif new file mode 100644 index 00000000..e6d0c8b9 Binary files /dev/null and b/assets/readme/sample_3.gif differ diff --git a/assets/readme/sample_4.gif b/assets/readme/sample_4.gif new file mode 100644 index 00000000..137a28e3 Binary files /dev/null and b/assets/readme/sample_4.gif differ diff --git a/assets/readme/sample_5.gif b/assets/readme/sample_5.gif new file mode 100644 index 00000000..375895b3 Binary files /dev/null and b/assets/readme/sample_5.gif differ diff --git a/assets/texts/imagenet_id.txt b/assets/texts/imagenet_id.txt new file mode 100644 index 00000000..9085aa00 --- /dev/null +++ b/assets/texts/imagenet_id.txt @@ -0,0 +1,8 @@ +207 +360 +387 +974 +88 +979 +417 +279 diff --git a/assets/texts/imagenet_labels.txt b/assets/texts/imagenet_labels.txt new file mode 100644 index 00000000..6493fdbf --- /dev/null +++ b/assets/texts/imagenet_labels.txt @@ -0,0 +1,8 @@ +golden retriever +otter +lesser panda +geyser +macaw +valley +balloon +golden panda diff --git a/assets/texts/t2i_samples.txt b/assets/texts/t2i_samples.txt new file mode 100644 index 00000000..9b729527 --- /dev/null +++ b/assets/texts/t2i_samples.txt @@ -0,0 +1,8 @@ +A small cactus with a happy face in the Sahara desert. +Bright scene, aerial view,ancient city, fantasy, gorgeous light, mirror reflection, high detail, wide angle lens. +Nature vs human nature, surreal, UHD, 8k, hyper details, rich colors, photograph. +Poster of a mechanical cat, techical Schematics viewed from front. +Luffy from ONEPIECE, handsome face, fantasy. +Real beautiful woman. +A alpaca made of colorful building blocks, cyberpunk. +artistic diff --git a/assets/texts/t2v_latte.txt b/assets/texts/t2v_latte.txt new file mode 100644 index 00000000..a61359ca --- /dev/null +++ b/assets/texts/t2v_latte.txt @@ -0,0 +1,7 @@ +Yellow and black tropical fish dart through the sea. +An epic tornado attacking above aglowing city at night. +Slow pan upward of blazing oak fire in an indoor fireplace. +a cat wearing sunglasses and working as a lifeguard at pool. +Sunset over the sea. +A dog in astronaut suit and sunglasses floating in space. +A astronaut in flying in space, 4k, high resolution diff --git a/assets/texts/t2v_samples.txt b/assets/texts/t2v_samples.txt new file mode 100644 index 00000000..312db460 --- /dev/null +++ b/assets/texts/t2v_samples.txt @@ -0,0 +1,10 @@ +A soaring drone footage captures the majestic beauty of a coastal cliff, its red and yellow stratified rock faces rich in color and against the vibrant turquoise of the sea. Seabirds can be seen taking flight around the cliff's precipices. As the drone slowly moves from different angles, the changing sunlight casts shifting shadows that highlight the rugged textures of the cliff and the surrounding calm sea. The water gently laps at the rock base and the greenery that clings to the top of the cliff, and the scene gives a sense of peaceful isolation at the fringes of the ocean. The video captures the essence of pristine natural beauty untouched by human structures. +The video captures the majestic beauty of a waterfall cascading down a cliff into a serene lake. The waterfall, with its powerful flow, is the central focus of the video. The surrounding landscape is lush and green, with trees and foliage adding to the natural beauty of the scene. The camera angle provides a bird's eye view of the waterfall, allowing viewers to appreciate the full height and grandeur of the waterfall. The video is a stunning representation of nature's power and beauty. +A vibrant scene of a snowy mountain landscape. The sky is filled with a multitude of colorful hot air balloons, each floating at different heights, creating a dynamic and lively atmosphere. The balloons are scattered across the sky, some closer to the viewer, others further away, adding depth to the scene. Below, the mountainous terrain is blanketed in a thick layer of snow, with a few patches of bare earth visible here and there. The snow-covered mountains provide a stark contrast to the colorful balloons, enhancing the visual appeal of the scene. In the foreground, a few cars can be seen driving along a winding road that cuts through the mountains. The cars are small compared to the vastness of the landscape, emphasizing the grandeur of the surroundings. The overall style of the video is a mix of adventure and tranquility, with the hot air balloons adding a touch of whimsy to the otherwise serene mountain landscape. The video is likely shot during the day, as the lighting is bright and even, casting soft shadows on the snow-covered mountains. +The vibrant beauty of a sunflower field. The sunflowers, with their bright yellow petals and dark brown centers, are in full bloom, creating a stunning contrast against the green leaves and stems. The sunflowers are arranged in neat rows, creating a sense of order and symmetry. The sun is shining brightly, casting a warm glow on the flowers and highlighting their intricate details. The video is shot from a low angle, looking up at the sunflowers, which adds a sense of grandeur and awe to the scene. The sunflowers are the main focus of the video, with no other objects or people present. The video is a celebration of nature's beauty and the simple joy of a sunny day in the countryside. +A serene underwater scene featuring a sea turtle swimming through a coral reef. The turtle, with its greenish-brown shell, is the main focus of the video, swimming gracefully towards the right side of the frame. The coral reef, teeming with life, is visible in the background, providing a vibrant and colorful backdrop to the turtle's journey. Several small fish, darting around the turtle, add a sense of movement and dynamism to the scene. The video is shot from a slightly elevated angle, providing a comprehensive view of the turtle's surroundings. The overall style of the video is calm and peaceful, capturing the beauty and tranquility of the underwater world. +A vibrant underwater scene. A group of blue fish, with yellow fins, are swimming around a coral reef. The coral reef is a mix of brown and green, providing a natural habitat for the fish. The water is a deep blue, indicating a depth of around 30 feet. The fish are swimming in a circular pattern around the coral reef, indicating a sense of motion and activity. The overall scene is a beautiful representation of marine life. +A bustling city street at night, filled with the glow of car headlights and the ambient light of streetlights. The scene is a blur of motion, with cars speeding by and pedestrians navigating the crosswalks. The cityscape is a mix of towering buildings and illuminated signs, creating a vibrant and dynamic atmosphere. The perspective of the video is from a high angle, providing a bird's eye view of the street and its surroundings. The overall style of the video is dynamic and energetic, capturing the essence of urban life at night. +A snowy forest landscape with a dirt road running through it. The road is flanked by trees covered in snow, and the ground is also covered in snow. The sun is shining, creating a bright and serene atmosphere. The road appears to be empty, and there are no people or animals visible in the video. The style of the video is a natural landscape shot, with a focus on the beauty of the snowy forest and the peacefulness of the road. +The dynamic movement of tall, wispy grasses swaying in the wind. The sky above is filled with clouds, creating a dramatic backdrop. The sunlight pierces through the clouds, casting a warm glow on the scene. The grasses are a mix of green and brown, indicating a change in seasons. The overall style of the video is naturalistic, capturing the beauty of the landscape in a realistic manner. The focus is on the grasses and their movement, with the sky serving as a secondary element. The video does not contain any human or animal elements. +A serene night scene in a forested area. The first frame shows a tranquil lake reflecting the star-filled sky above. The second frame reveals a beautiful sunset, casting a warm glow over the landscape. The third frame showcases the night sky, filled with stars and a vibrant Milky Way galaxy. The video is a time-lapse, capturing the transition from day to night, with the lake and forest serving as a constant backdrop. The style of the video is naturalistic, emphasizing the beauty of the night sky and the peacefulness of the forest. diff --git a/assets/texts/t2v_sora.txt b/assets/texts/t2v_sora.txt new file mode 100644 index 00000000..eeb887b1 --- /dev/null +++ b/assets/texts/t2v_sora.txt @@ -0,0 +1,48 @@ +A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. She wears a black leather jacket, a long red dress, and black boots, and carries a black purse. She wears sunglasses and red lipstick. She walks confidently and casually. The street is damp and reflective, creating a mirror effect of the colorful lights. Many pedestrians walk about. +Several giant wooly mammoths approach treading through a snowy meadow, their long wooly fur lightly blows in the wind as they walk, snow covered trees and dramatic snow capped mountains in the distance, mid afternoon light with wispy clouds and a sun high in the distance creates a warm glow, the low camera view is stunning capturing the large furry mammal with beautiful photography, depth of field. +A movie trailer featuring the adventures of the 30 year old space man wearing a red wool knitted motorcycle helmet, blue sky, salt desert, cinematic style, shot on 35mm film, vivid colors. +Drone view of waves crashing against the rugged cliffs along Big Sur’s garay point beach. The crashing blue waters create white-tipped waves, while the golden light of the setting sun illuminates the rocky shore. A small island with a lighthouse sits in the distance, and green shrubbery covers the cliff’s edge. The steep drop from the road down to the beach is a dramatic feat, with the cliff’s edges jutting out over the sea. This is a view that captures the raw beauty of the coast and the rugged landscape of the Pacific Coast Highway. +Animated scene features a close-up of a short fluffy monster kneeling beside a melting red candle. The art style is 3D and realistic, with a focus on lighting and texture. The mood of the painting is one of wonder and curiosity, as the monster gazes at the flame with wide eyes and open mouth. Its pose and expression convey a sense of innocence and playfulness, as if it is exploring the world around it for the first time. The use of warm colors and dramatic lighting further enhances the cozy atmosphere of the image. +A gorgeously rendered papercraft world of a coral reef, rife with colorful fish and sea creatures. +This close-up shot of a Victoria crowned pigeon showcases its striking blue plumage and red chest. Its crest is made of delicate, lacy feathers, while its eye is a striking red color. The bird’s head is tilted slightly to the side, giving the impression of it looking regal and majestic. The background is blurred, drawing attention to the bird’s striking appearance. +Photorealistic closeup video of two pirate ships battling each other as they sail inside a cup of coffee. +A young man at his 20s is sitting on a piece of cloud in the sky, reading a book. +Historical footage of California during the gold rush. +A close up view of a glass sphere that has a zen garden within it. There is a small dwarf in the sphere who is raking the zen garden and creating patterns in the sand. +Extreme close up of a 24 year old woman’s eye blinking, standing in Marrakech during magic hour, cinematic film shot in 70mm, depth of field, vivid colors, cinematic +A cartoon kangaroo disco dances. +A beautiful homemade video showing the people of Lagos, Nigeria in the year 2056. Shot with a mobile phone camera. +A petri dish with a bamboo forest growing within it that has tiny red pandas running around. +The camera rotates around a large stack of vintage televisions all showing different programs — 1950s sci-fi movies, horror movies, news, static, a 1970s sitcom, etc, set inside a large New York museum gallery. +3D animation of a small, round, fluffy creature with big, expressive eyes explores a vibrant, enchanted forest. The creature, a whimsical blend of a rabbit and a squirrel, has soft blue fur and a bushy, striped tail. It hops along a sparkling stream, its eyes wide with wonder. The forest is alive with magical elements: flowers that glow and change colors, trees with leaves in shades of purple and silver, and small floating lights that resemble fireflies. The creature stops to interact playfully with a group of tiny, fairy-like beings dancing around a mushroom ring. The creature looks up in awe at a large, glowing tree that seems to be the heart of the forest. +The camera follows behind a white vintage SUV with a black roof rack as it speeds up a steep dirt road surrounded by pine trees on a steep mountain slope, dust kicks up from it’s tires, the sunlight shines on the SUV as it speeds along the dirt road, casting a warm glow over the scene. The dirt road curves gently into the distance, with no other cars or vehicles in sight. The trees on either side of the road are redwoods, with patches of greenery scattered throughout. The car is seen from the rear following the curve with ease, making it seem as if it is on a rugged drive through the rugged terrain. The dirt road itself is surrounded by steep hills and mountains, with a clear blue sky above with wispy clouds. +Reflections in the window of a train traveling through the Tokyo suburbs. +A drone camera circles around a beautiful historic church built on a rocky outcropping along the Amalfi Coast, the view showcases historic and magnificent architectural details and tiered pathways and patios, waves are seen crashing against the rocks below as the view overlooks the horizon of the coastal waters and hilly landscapes of the Amalfi Coast Italy, several distant people are seen walking and enjoying vistas on patios of the dramatic ocean views, the warm glow of the afternoon sun creates a magical and romantic feeling to the scene, the view is stunning captured with beautiful photography. +A large orange octopus is seen resting on the bottom of the ocean floor, blending in with the sandy and rocky terrain. Its tentacles are spread out around its body, and its eyes are closed. The octopus is unaware of a king crab that is crawling towards it from behind a rock, its claws raised and ready to attack. The crab is brown and spiny, with long legs and antennae. The scene is captured from a wide angle, showing the vastness and depth of the ocean. The water is clear and blue, with rays of sunlight filtering through. The shot is sharp and crisp, with a high dynamic range. The octopus and the crab are in focus, while the background is slightly blurred, creating a depth of field effect. +A flock of paper airplanes flutters through a dense jungle, weaving around trees as if they were migrating birds. +A cat waking up its sleeping owner demanding breakfast. The owner tries to ignore the cat, but the cat tries new tactics and finally the owner pulls out a secret stash of treats from under the pillow to hold the cat off a little longer. +Borneo wildlife on the Kinabatangan River +A Chinese Lunar New Year celebration video with Chinese Dragon. +Tour of an art gallery with many beautiful works of art in different styles. +Beautiful, snowy Tokyo city is bustling. The camera moves through the bustling city street, following several people enjoying the beautiful snowy weather and shopping at nearby stalls. Gorgeous sakura petals are flying through the wind along with snowflakes. +A stop motion animation of a flower growing out of the windowsill of a suburban house. +The story of a robot’s life in a cyberpunk setting. +An extreme close-up of an gray-haired man with a beard in his 60s, he is deep in thought pondering the history of the universe as he sits at a cafe in Paris, his eyes focus on people offscreen as they walk as he sits mostly motionless, he is dressed in a wool coat suit coat with a button-down shirt , he wears a brown beret and glasses and has a very professorial appearance, and the end he offers a subtle closed-mouth smile as if he found the answer to the mystery of life, the lighting is very cinematic with the golden light and the Parisian streets and city in the background, depth of field, cinematic 35mm film. +A beautiful silhouette animation shows a wolf howling at the moon, feeling lonely, until it finds its pack. +New York City submerged like Atlantis. Fish, whales, sea turtles and sharks swim through the streets of New York. +A litter of golden retriever puppies playing in the snow. Their heads pop out of the snow, covered in. +Step-printing scene of a person running, cinematic film shot in 35mm. +Five gray wolf pups frolicking and chasing each other around a remote gravel road, surrounded by grass. The pups run and leap, chasing each other, and nipping at each other, playing. +Basketball through hoop then explodes. +Archeologists discover a generic plastic chair in the desert, excavating and dusting it with great care. +A grandmother with neatly combed grey hair stands behind a colorful birthday cake with numerous candles at a wood dining room table, expression is one of pure joy and happiness, with a happy glow in her eye. She leans forward and blows out the candles with a gentle puff, the cake has pink frosting and sprinkles and the candles cease to flicker, the grandmother wears a light blue blouse adorned with floral patterns, several happy friends and family sitting at the table can be seen celebrating, out of focus. The scene is beautifully captured, cinematic, showing a 3/4 view of the grandmother and the dining room. Warm color tones and soft lighting enhance the mood. +The camera directly faces colorful buildings in Burano Italy. An adorable dalmation looks through a window on a building on the ground floor. Many people are walking and cycling along the canal streets in front of the buildings. +An adorable happy otter confidently stands on a surfboard wearing a yellow lifejacket, riding along turquoise tropical waters near lush tropical islands, 3D digital render art style. +This close-up shot of a chameleon showcases its striking color changing capabilities. The background is blurred, drawing attention to the animal’s striking appearance. +A corgi vlogging itself in tropical Maui. +A white and orange tabby cat is seen happily darting through a dense garden, as if chasing something. Its eyes are wide and happy as it jogs forward, scanning the branches, flowers, and leaves as it walks. The path is narrow as it makes its way between all the plants. the scene is captured from a ground-level angle, following the cat closely, giving a low and intimate perspective. The image is cinematic with warm tones and a grainy texture. The scattered daylight between the leaves and plants above creates a warm contrast, accentuating the cat’s orange fur. The shot is clear and sharp, with a shallow depth of field. +Aerial view of Santorini during the blue hour, showcasing the stunning architecture of white Cycladic buildings with blue domes. The caldera views are breathtaking, and the lighting creates a beautiful, serene atmosphere. +Tiltshift of a construction site filled with workers, equipment, and heavy machinery. +A giant, towering cloud in the shape of a man looms over the earth. The cloud man shoots lighting bolts down to the earth. +A Samoyed and a Golden Retriever dog are playfully romping through a futuristic neon city at night. The neon lights emitted from the nearby buildings glistens off of their fur. +The Glenfinnan Viaduct is a historic railway bridge in Scotland, UK, that crosses over the west highland line between the towns of Mallaig and Fort William. It is a stunning sight as a steam train leaves the bridge, traveling over the arch-covered viaduct. The landscape is dotted with lush greenery and rocky mountains, creating a picturesque backdrop for the train journey. The sky is blue and the sun is shining, making for a beautiful day to explore this majestic spot. diff --git a/assets/texts/ucf101_id.txt b/assets/texts/ucf101_id.txt new file mode 100644 index 00000000..e8371f00 --- /dev/null +++ b/assets/texts/ucf101_id.txt @@ -0,0 +1,6 @@ +0 +1 +2 +3 +4 +5 diff --git a/assets/texts/ucf101_labels.txt b/assets/texts/ucf101_labels.txt new file mode 100644 index 00000000..264dbfd8 --- /dev/null +++ b/assets/texts/ucf101_labels.txt @@ -0,0 +1,6 @@ +Apply Eye Makeup +Apply Lipstick +Archery +Baby Crawling +Balance Beam +Band Marching diff --git a/benchmark.py b/benchmark.py deleted file mode 100644 index 69eef8d2..00000000 --- a/benchmark.py +++ /dev/null @@ -1,180 +0,0 @@ -# Copyright (c) Meta Platforms, Inc. and affiliates. -# All rights reserved. - -# This source code is licensed under the license found in the -# LICENSE file in the root directory of this source tree. - -""" -A minimal training script for DiT using PyTorch DDP. -""" -import argparse -import time - -import torch -import torch.distributed as dist -from colossalai import launch_from_torch -from colossalai.accelerator import get_accelerator -from colossalai.booster import Booster -from colossalai.booster.plugin import TorchDDPPlugin -from colossalai.cluster import DistCoordinator -from colossalai.logging import get_dist_logger -from colossalai.nn.optimizer import HybridAdam -from colossalai.utils import get_current_device -from tqdm import tqdm - -from open_sora.diffusion import create_diffusion -from open_sora.modeling import DiT_models -from open_sora.modeling.dit import SUPPORTED_SEQ_PARALLEL_MODES -from open_sora.utils.data import create_video_compressor, preprocess_batch -from open_sora.utils.plugin import ZeroSeqParallelPlugin - -################################################################################# -# Training Loop # -################################################################################# - - -def main(args): - """ - Trains a new DiT model. - """ - # init distributed environment - launch_from_torch({}) - coordinator = DistCoordinator() - logger = get_dist_logger() - - # set up acceleration plugins - if args.plugin == "ddp": - plugin = TorchDDPPlugin() - elif args.plugin == "zero2": - # use bf16 to avoid skipping the first few iterations due to NaNs - plugin = ZeroSeqParallelPlugin(sp_size=args.sp_size, stage=2, precision="bf16") - else: - raise ValueError(f"Unknown plugin {args.plugin}") - booster = Booster(plugin=plugin) - - # Create video compressor - video_compressor = create_video_compressor(args.compressor) - model_kwargs = { - "in_channels": video_compressor.out_channels, - "seq_parallel_group": getattr(plugin, "sp_group", None), - "seq_parallel_mode": args.sp_mode, - "seq_parallel_overlap": args.sp_overlap, - } - - # Create DiT and EMA - model = DiT_models[args.model](**model_kwargs).to(get_current_device()) - patch_size = model.patch_size - model.train() # important! This enables embedding dropout for classifier-free guidance - - # configure gradient checkpointing - if args.grad_checkpoint: - model.enable_gradient_checkpointing() - - # create diffusion pipeline - diffusion = create_diffusion( - timestep_respacing="" - ) # default: 1000 steps, linear noise schedule - - # setup optimizer (we used default Adam betas=(0.9, 0.999) and a constant learning rate of 1e-4 in our paper): - opt = HybridAdam(model.parameters(), lr=1e-4, weight_decay=0) - - # Setup dataloader - videos = [ - torch.randn(args.num_frames, args.height, args.width, 3) - for _ in range(args.batch_size) - ] - assert args.num_tokens % args.sp_size == 0 - input_ids = torch.randn(args.batch_size, args.num_tokens, args.text_embed_dim) - text_mask = torch.ones(input_ids.shape[:2], dtype=torch.int) - batch = { - "videos": videos, - "text_latent_states": input_ids, - "text_padding_mask": text_mask, - } - batch = preprocess_batch( - batch, patch_size, video_compressor, pad_to_multiple=args.sp_size - ) - video_inputs = batch.pop("video_latent_states") - mask = batch.pop("video_padding_mask") - logger.info( - f"Num patches: {video_inputs.shape[1]}, num tokens: {batch['text_latent_states'].shape[1]}", - ranks=[0], - ) - - # setup booster - model, opt, *_ = booster.boost(model, opt) - logger.info( - f"Booster init max device memory: {get_accelerator().max_memory_allocated() / 1024 ** 2:.2f} MB", - ranks=[0], - ) - - # Train - total_samples = 0 - total_duration = 0.0 - for i in tqdm( - range(args.warmup_steps + args.steps), - desc="Steps", - disable=not coordinator.is_master(), - ): - start = time.time() - t = torch.randint( - 0, - diffusion.num_timesteps, - (video_inputs.shape[0],), - device=video_inputs.device, - ) - loss_dict = diffusion.training_losses(model, video_inputs, t, batch, mask=mask) - loss = loss_dict["loss"].mean() - booster.backward(loss, opt) - opt.step() - opt.zero_grad() - get_accelerator().empty_cache() - time_per_iter = time.time() - start - if i >= args.warmup_steps: - total_samples += args.batch_size * coordinator.world_size - total_duration += time_per_iter - total_duration = torch.tensor([total_duration], device=get_current_device()) - dist.all_reduce(total_duration) - total_duration = total_duration / coordinator.world_size - total_duration = total_duration.item() - total_samples *= coordinator.world_size // args.sp_size - - throughput = total_samples / total_duration - logger.info( - f"Training complete, max device memory: {get_accelerator().max_memory_allocated() / 1024 ** 2:.2f} MB", - ranks=[0], - ) - logger.info( - f"Throughput: {throughput:.2f} samples/s", - ranks=[0], - ) - - -if __name__ == "__main__": - # Default args here will train DiT-XL/2 with the hyperparameters we used in our paper (except training iters). - parser = argparse.ArgumentParser() - parser.add_argument( - "-m", "--model", type=str, choices=list(DiT_models.keys()), default="DiT-S/8" - ) - parser.add_argument( - "-p", "--plugin", type=str, default="zero2", choices=["ddp", "zero2"] - ) - parser.add_argument("--sp_size", type=int, default=1) - parser.add_argument( - "--sp_mode", type=str, default="ulysses", choices=SUPPORTED_SEQ_PARALLEL_MODES - ) - parser.add_argument("--sp_overlap", action="store_true", default=False) - parser.add_argument("-w", "--warmup_steps", type=int, default=2) - parser.add_argument("-s", "--steps", type=int, default=3) - parser.add_argument("-b", "--batch_size", type=int, default=4) - parser.add_argument("-f", "--num_frames", type=int, default=300) - parser.add_argument("--height", type=int, default=256) - parser.add_argument("--width", type=int, default=256) - parser.add_argument("--num_tokens", type=int, default=20) - parser.add_argument("--text_embed_dim", type=int, default=512) - parser.add_argument("-g", "--grad_checkpoint", action="store_true", default=False) - parser.add_argument( - "-c", "--compressor", choices=["raw", "vqvae", "vae"], default="raw" - ) - args = parser.parse_args() - main(args) diff --git a/configs/dit/inference/16x256x256.py b/configs/dit/inference/16x256x256.py new file mode 100644 index 00000000..ccb1d796 --- /dev/null +++ b/configs/dit/inference/16x256x256.py @@ -0,0 +1,31 @@ +num_frames = 16 +fps = 8 +image_size = (256, 256) + +# Define model +model = dict( + type="DiT-XL/2", + condition="text", + from_pretrained="PRETRAINED_MODEL", +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", +) +text_encoder = dict( + type="clip", + from_pretrained="openai/clip-vit-base-patch32", + model_max_length=77, +) +scheduler = dict( + type="dpm-solver", + num_sampling_steps=20, + cfg_scale=4.0, +) +dtype = "fp16" + +# Others +batch_size = 2 +seed = 42 +prompt_path = "./assets/texts/ucf101_labels.txt" +save_dir = "./outputs/samples/" diff --git a/configs/dit/inference/1x256x256-class.py b/configs/dit/inference/1x256x256-class.py new file mode 100644 index 00000000..24d1c8af --- /dev/null +++ b/configs/dit/inference/1x256x256-class.py @@ -0,0 +1,31 @@ +num_frames = 1 +fps = 1 +image_size = (256, 256) + +# Define model +model = dict( + type="DiT-XL/2", + no_temporal_pos_emb=True, + condition="label_1000", + from_pretrained="DiT-XL-2-256x256.pt", +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", +) +text_encoder = dict( + type="classes", + num_classes=1000, +) +scheduler = dict( + type="dpm-solver", + num_sampling_steps=20, + cfg_scale=4.0, +) +dtype = "fp16" + +# Others +batch_size = 2 +seed = 42 +prompt_path = "./assets/texts/imagenet_id.txt" +save_dir = "./outputs/samples/" diff --git a/configs/dit/inference/1x256x256.py b/configs/dit/inference/1x256x256.py new file mode 100644 index 00000000..31a5b9f1 --- /dev/null +++ b/configs/dit/inference/1x256x256.py @@ -0,0 +1,32 @@ +num_frames = 1 +fps = 1 +image_size = (256, 256) + +# Define model +model = dict( + type="DiT-XL/2", + no_temporal_pos_emb=True, + condition="text", + from_pretrained="PRETRAINED_MODEL", +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", +) +text_encoder = dict( + type="clip", + from_pretrained="openai/clip-vit-base-patch32", + model_max_length=77, +) +scheduler = dict( + type="dpm-solver", + num_sampling_steps=20, + cfg_scale=4.0, +) +dtype = "fp16" + +# Others +batch_size = 2 +seed = 42 +prompt_path = "./assets/texts/imagenet_labels.txt" +save_dir = "./outputs/samples/" diff --git a/configs/dit/train/16x256x256.py b/configs/dit/train/16x256x256.py new file mode 100644 index 00000000..af8ee876 --- /dev/null +++ b/configs/dit/train/16x256x256.py @@ -0,0 +1,50 @@ +num_frames = 16 +frame_interval = 3 +image_size = (256, 256) + +# Define dataset +root = None +data_path = "CSV_PATH" +use_image_transform = False +num_workers = 4 + +# Define acceleration +dtype = "bf16" +grad_checkpoint = False +plugin = "zero2" +sp_size = 1 + +# Define model +model = dict( + type="DiT-XL/2", + from_pretrained="DiT-XL-2-256x256.pt", + enable_flashattn=True, + enable_layernorm_kernel=True, +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", +) +text_encoder = dict( + type="clip", + from_pretrained="openai/clip-vit-base-patch32", + model_max_length=77, +) +scheduler = dict( + type="iddpm", + timestep_respacing="", +) + +# Others +seed = 42 +outputs = "outputs" +wandb = False + +epochs = 1000 +log_every = 10 +ckpt_every = 1000 +load = None + +batch_size = 8 +lr = 2e-5 +grad_clip = 1.0 diff --git a/configs/dit/train/1x256x256.py b/configs/dit/train/1x256x256.py new file mode 100644 index 00000000..667e0a83 --- /dev/null +++ b/configs/dit/train/1x256x256.py @@ -0,0 +1,50 @@ +num_frames = 1 +frame_interval = 1 +image_size = (256, 256) + +# Define dataset +root = None +data_path = "CSV_PATH" +use_image_transform = True +num_workers = 4 + +# Define acceleration +dtype = "bf16" +grad_checkpoint = False +plugin = "zero2" +sp_size = 1 + +# Define model +model = dict( + type="DiT-XL/2", + no_temporal_pos_emb=True, + enable_flashattn=True, + enable_layernorm_kernel=True, +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", +) +text_encoder = dict( + type="clip", + from_pretrained="openai/clip-vit-base-patch32", + model_max_length=77, +) +scheduler = dict( + type="iddpm", + timestep_respacing="", +) + +# Others +seed = 42 +outputs = "outputs" +wandb = False + +epochs = 1000 +log_every = 10 +ckpt_every = 1000 +load = None + +batch_size = 128 +lr = 1e-4 # according to DiT repo +grad_clip = 1.0 diff --git a/configs/latte/inference/16x256x256-class.py b/configs/latte/inference/16x256x256-class.py new file mode 100644 index 00000000..c46f4bc3 --- /dev/null +++ b/configs/latte/inference/16x256x256-class.py @@ -0,0 +1,30 @@ +num_frames = 16 +fps = 8 +image_size = (256, 256) + +# Define model +model = dict( + type="Latte-XL/2", + condition="label_101", + from_pretrained="Latte-XL-2-256x256-ucf101.pt", +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", +) +text_encoder = dict( + type="classes", + num_classes=101, +) +scheduler = dict( + type="dpm-solver", + num_sampling_steps=20, + cfg_scale=4.0, +) +dtype = "fp16" + +# Others +batch_size = 2 +seed = 42 +prompt_path = "./assets/texts/ucf101_id.txt" +save_dir = "./outputs/samples/" diff --git a/configs/latte/inference/16x256x256.py b/configs/latte/inference/16x256x256.py new file mode 100644 index 00000000..cb502371 --- /dev/null +++ b/configs/latte/inference/16x256x256.py @@ -0,0 +1,31 @@ +num_frames = 16 +fps = 8 +image_size = (256, 256) + +# Define model +model = dict( + type="Latte-XL/2", + condition="text", + from_pretrained="PRETRAINED_MODEL", +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", +) +text_encoder = dict( + type="clip", + from_pretrained="openai/clip-vit-base-patch32", + model_max_length=77, +) +scheduler = dict( + type="dpm-solver", + num_sampling_steps=20, + cfg_scale=4.0, +) +dtype = "fp16" + +# Others +batch_size = 2 +seed = 42 +prompt_path = "./assets/texts/ucf101_labels.txt" +save_dir = "./outputs/samples/" diff --git a/configs/latte/train/16x256x256.py b/configs/latte/train/16x256x256.py new file mode 100644 index 00000000..0bf6bd41 --- /dev/null +++ b/configs/latte/train/16x256x256.py @@ -0,0 +1,49 @@ +num_frames = 16 +frame_interval = 3 +image_size = (256, 256) + +# Define dataset +root = None +data_path = "CSV_PATH" +use_image_transform = False +num_workers = 4 + +# Define acceleration +dtype = "bf16" +grad_checkpoint = True +plugin = "zero2" +sp_size = 1 + +# Define model +model = dict( + type="Latte-XL/2", + enable_flashattn=True, + enable_layernorm_kernel=True, +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", +) +text_encoder = dict( + type="clip", + from_pretrained="openai/clip-vit-base-patch32", + model_max_length=77, +) +scheduler = dict( + type="iddpm", + timestep_respacing="", +) + +# Others +seed = 42 +outputs = "outputs" +wandb = False + +epochs = 1000 +log_every = 10 +ckpt_every = 1000 +load = None + +batch_size = 8 +lr = 2e-5 +grad_clip = 1.0 diff --git a/configs/opensora/inference/16x256x256.py b/configs/opensora/inference/16x256x256.py new file mode 100644 index 00000000..5e1a47e5 --- /dev/null +++ b/configs/opensora/inference/16x256x256.py @@ -0,0 +1,34 @@ +num_frames = 16 +fps = 24 // 3 +image_size = (256, 256) + +# Define model +model = dict( + type="STDiT-XL/2", + space_scale=0.5, + time_scale=1.0, + enable_flashattn=True, + enable_layernorm_kernel=True, + from_pretrained="PRETRAINED_MODEL", +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", +) +text_encoder = dict( + type="t5", + from_pretrained="./pretrained_models/t5_ckpts", + model_max_length=120, +) +scheduler = dict( + type="iddpm", + num_sampling_steps=100, + cfg_scale=7.0, +) +dtype = "fp16" + +# Others +batch_size = 2 +seed = 42 +prompt_path = "./assets/texts/t2v_samples.txt" +save_dir = "./outputs/samples/" diff --git a/configs/opensora/inference/16x512x512.py b/configs/opensora/inference/16x512x512.py new file mode 100644 index 00000000..b64b85c7 --- /dev/null +++ b/configs/opensora/inference/16x512x512.py @@ -0,0 +1,35 @@ +num_frames = 16 +fps = 24 // 3 +image_size = (512, 512) + +# Define model +model = dict( + type="STDiT-XL/2", + space_scale=1.0, + time_scale=1.0, + enable_flashattn=True, + enable_layernorm_kernel=True, + from_pretrained="PRETRAINED_MODEL" +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", + micro_batch_size=128, +) +text_encoder = dict( + type="t5", + from_pretrained="./pretrained_models/t5_ckpts", + model_max_length=120, +) +scheduler = dict( + type="iddpm", + num_sampling_steps=100, + cfg_scale=7.0, +) +dtype = "fp16" + +# Others +batch_size = 2 +seed = 42 +prompt_path = "./assets/texts/t2v_samples.txt" +save_dir = "./outputs/samples/" diff --git a/configs/opensora/inference/64x512x512.py b/configs/opensora/inference/64x512x512.py new file mode 100644 index 00000000..e15649a3 --- /dev/null +++ b/configs/opensora/inference/64x512x512.py @@ -0,0 +1,35 @@ +num_frames = 64 +fps = 24 // 2 +image_size = (512, 512) + +# Define model +model = dict( + type="STDiT-XL/2", + space_scale=1.0, + time_scale=2 / 3, + enable_flashattn=True, + enable_layernorm_kernel=True, + from_pretrained="PRETRAINED_MODEL", +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", + micro_batch_size=128, +) +text_encoder = dict( + type="t5", + from_pretrained="./pretrained_models/t5_ckpts", + model_max_length=120, +) +scheduler = dict( + type="iddpm", + num_sampling_steps=100, + cfg_scale=7.0, +) +dtype = "fp16" + +# Others +batch_size = 1 +seed = 42 +prompt_path = "./assets/texts/t2v_samples.txt" +save_dir = "./outputs/samples/" diff --git a/configs/opensora/train/16x256x256.py b/configs/opensora/train/16x256x256.py new file mode 100644 index 00000000..a64a318f --- /dev/null +++ b/configs/opensora/train/16x256x256.py @@ -0,0 +1,53 @@ +num_frames = 16 +frame_interval = 3 +image_size = (256, 256) + +# Define dataset +root = None +data_path = "CSV_PATH" +use_image_transform = False +num_workers = 4 + +# Define acceleration +dtype = "bf16" +grad_checkpoint = True +plugin = "zero2" +sp_size = 1 + +# Define model +model = dict( + type="STDiT-XL/2", + space_scale=0.5, + time_scale=1.0, + from_pretrained="PixArt-XL-2-512x512.pth", + enable_flashattn=True, + enable_layernorm_kernel=True, +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", +) +text_encoder = dict( + type="t5", + from_pretrained="./pretrained_models/t5_ckpts", + model_max_length=120, + shardformer=True, +) +scheduler = dict( + type="iddpm", + timestep_respacing="", +) + +# Others +seed = 42 +outputs = "outputs" +wandb = False + +epochs = 1000 +log_every = 10 +ckpt_every = 1000 +load = None + +batch_size = 8 +lr = 2e-5 +grad_clip = 1.0 diff --git a/configs/opensora/train/16x512x512.py b/configs/opensora/train/16x512x512.py new file mode 100644 index 00000000..885aad1f --- /dev/null +++ b/configs/opensora/train/16x512x512.py @@ -0,0 +1,54 @@ +num_frames = 16 +frame_interval = 3 +image_size = (512, 512) + +# Define dataset +root = None +data_path = "CSV_PATH" +use_image_transform = False +num_workers = 4 + +# Define acceleration +dtype = "bf16" +grad_checkpoint = False +plugin = "zero2" +sp_size = 1 + +# Define model +model = dict( + type="STDiT-XL/2", + space_scale=1.0, + time_scale=1.0, + from_pretrained=None, + enable_flashattn=True, + enable_layernorm_kernel=True, +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", + micro_batch_size=128, +) +text_encoder = dict( + type="t5", + from_pretrained="./pretrained_models/t5_ckpts", + model_max_length=120, + shardformer=True, +) +scheduler = dict( + type="iddpm", + timestep_respacing="", +) + +# Others +seed = 42 +outputs = "outputs" +wandb = False + +epochs = 1000 +log_every = 10 +ckpt_every = 500 +load = None + +batch_size = 8 +lr = 2e-5 +grad_clip = 1.0 diff --git a/configs/opensora/train/360x512x512.py b/configs/opensora/train/360x512x512.py new file mode 100644 index 00000000..7a6f7599 --- /dev/null +++ b/configs/opensora/train/360x512x512.py @@ -0,0 +1,55 @@ +num_frames = 360 +frame_interval = 1 +image_size = (512, 512) + +# Define dataset +root = None +data_path = "CSV_PATH" +use_image_transform = False +num_workers = 4 + +# Define acceleration +dtype = "bf16" +grad_checkpoint = True +plugin = "zero2-seq" +sp_size = 2 + +# Define model +model = dict( + type="STDiT-XL/2", + space_scale=1.0, + time_scale=2 / 3, + from_pretrained=None, + enable_flashattn=True, + enable_layernorm_kernel=True, + enable_sequence_parallelism=True, # enable sq here +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", + micro_batch_size=128, +) +text_encoder = dict( + type="t5", + from_pretrained="./pretrained_models/t5_ckpts", + model_max_length=120, + shardformer=True, +) +scheduler = dict( + type="iddpm", + timestep_respacing="", +) + +# Others +seed = 42 +outputs = "outputs" +wandb = False + +epochs = 1000 +log_every = 10 +ckpt_every = 250 +load = None + +batch_size = 1 +lr = 2e-5 +grad_clip = 1.0 diff --git a/configs/opensora/train/64x512x512-sp.py b/configs/opensora/train/64x512x512-sp.py new file mode 100644 index 00000000..b0b9062c --- /dev/null +++ b/configs/opensora/train/64x512x512-sp.py @@ -0,0 +1,54 @@ +num_frames = 64 +frame_interval = 2 +image_size = (512, 512) + +# Define dataset +root = None +data_path = "CSV_PATH" +use_image_transform = False +num_workers = 4 + +# Define acceleration +dtype = "bf16" +grad_checkpoint = True +plugin = "zero2-seq" +sp_size = 2 + +# Define model +model = dict( + type="STDiT-XL/2", + space_scale=1.0, + time_scale=2 / 3, + from_pretrained=None, + enable_flashattn=True, + enable_layernorm_kernel=True, + enable_sequence_parallelism=True, # enable sq here +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", +) +text_encoder = dict( + type="t5", + from_pretrained="./pretrained_models/t5_ckpts", + model_max_length=120, + shardformer=True, +) +scheduler = dict( + type="iddpm", + timestep_respacing="", +) + +# Others +seed = 42 +outputs = "outputs" +wandb = False + +epochs = 1000 +log_every = 10 +ckpt_every = 1000 +load = None + +batch_size = 1 +lr = 2e-5 +grad_clip = 1.0 diff --git a/configs/opensora/train/64x512x512.py b/configs/opensora/train/64x512x512.py new file mode 100644 index 00000000..dfcdcc08 --- /dev/null +++ b/configs/opensora/train/64x512x512.py @@ -0,0 +1,54 @@ +num_frames = 64 +frame_interval = 2 +image_size = (512, 512) + +# Define dataset +root = None +data_path = "CSV_PATH" +use_image_transform = False +num_workers = 4 + +# Define acceleration +dtype = "bf16" +grad_checkpoint = True +plugin = "zero2" +sp_size = 1 + +# Define model +model = dict( + type="STDiT-XL/2", + space_scale=1.0, + time_scale=2 / 3, + from_pretrained=None, + enable_flashattn=True, + enable_layernorm_kernel=True, +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", + micro_batch_size=64, +) +text_encoder = dict( + type="t5", + from_pretrained="./pretrained_models/t5_ckpts", + model_max_length=120, + shardformer=True, +) +scheduler = dict( + type="iddpm", + timestep_respacing="", +) + +# Others +seed = 42 +outputs = "outputs" +wandb = False + +epochs = 1000 +log_every = 10 +ckpt_every = 250 +load = None + +batch_size = 4 +lr = 2e-5 +grad_clip = 1.0 diff --git a/configs/pixart/inference/16x256x256.py b/configs/pixart/inference/16x256x256.py new file mode 100644 index 00000000..6fc8ee65 --- /dev/null +++ b/configs/pixart/inference/16x256x256.py @@ -0,0 +1,32 @@ +num_frames = 16 +fps = 8 +image_size = (256, 256) + +# Define model +model = dict( + type="PixArt-XL/2", + space_scale=0.5, + time_scale=1.0, + from_pretrained="outputs/098-F16S3-PixArt-XL-2/epoch7-global_step30000/model_ckpt.pt", +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", +) +text_encoder = dict( + type="t5", + from_pretrained="./pretrained_models/t5_ckpts", + model_max_length=120, +) +scheduler = dict( + type="dpm-solver", + num_sampling_steps=20, + cfg_scale=7.0, +) +dtype = "fp16" + +# Others +batch_size = 2 +seed = 42 +prompt_path = "./assets/texts/t2v_samples.txt" +save_dir = "./outputs/samples/" diff --git a/configs/pixart/inference/1x1024MS.py b/configs/pixart/inference/1x1024MS.py new file mode 100644 index 00000000..41cc97ad --- /dev/null +++ b/configs/pixart/inference/1x1024MS.py @@ -0,0 +1,34 @@ +num_frames = 1 +fps = 1 +image_size = (1920, 512) +multi_resolution = True + +# Define model +model = dict( + type="PixArtMS-XL/2", + space_scale=2.0, + time_scale=1.0, + no_temporal_pos_emb=True, + from_pretrained="PixArt-XL-2-1024-MS.pth", +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", +) +text_encoder = dict( + type="t5", + from_pretrained="./pretrained_models/t5_ckpts", + model_max_length=120, +) +scheduler = dict( + type="dpm-solver", + num_sampling_steps=20, + cfg_scale=7.0, +) +dtype = "fp16" + +# Others +batch_size = 2 +seed = 42 +prompt_path = "./assets/texts/t2i_samples.txt" +save_dir = "./outputs/samples/" diff --git a/configs/pixart/inference/1x256x256.py b/configs/pixart/inference/1x256x256.py new file mode 100644 index 00000000..11e06d77 --- /dev/null +++ b/configs/pixart/inference/1x256x256.py @@ -0,0 +1,33 @@ +num_frames = 1 +fps = 1 +image_size = (256, 256) + +# Define model +model = dict( + type="PixArt-XL/2", + space_scale=1.0, + time_scale=1.0, + no_temporal_pos_emb=True, + from_pretrained="PixArt-XL-2-256x256.pth", +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", +) +text_encoder = dict( + type="t5", + from_pretrained="./pretrained_models/t5_ckpts", + model_max_length=120, +) +scheduler = dict( + type="dpm-solver", + num_sampling_steps=20, + cfg_scale=7.0, +) +dtype = "fp16" + +# Others +batch_size = 2 +seed = 42 +prompt_path = "./assets/texts/t2i_samples.txt" +save_dir = "./outputs/samples/" diff --git a/configs/pixart/inference/1x512x512.py b/configs/pixart/inference/1x512x512.py new file mode 100644 index 00000000..5674259b --- /dev/null +++ b/configs/pixart/inference/1x512x512.py @@ -0,0 +1,33 @@ +num_frames = 1 +fps = 1 +image_size = (512, 512) + +# Define model +model = dict( + type="PixArt-XL/2", + space_scale=1.0, + time_scale=1.0, + no_temporal_pos_emb=True, + from_pretrained="PixArt-XL-2-512x512.pth", +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", +) +text_encoder = dict( + type="t5", + from_pretrained="./pretrained_models/t5_ckpts", + model_max_length=120, +) +scheduler = dict( + type="dpm-solver", + num_sampling_steps=20, + cfg_scale=7.0, +) +dtype = "fp16" + +# Others +batch_size = 2 +seed = 42 +prompt_path = "./assets/texts/t2i_samples.txt" +save_dir = "./outputs/samples/" diff --git a/configs/pixart/train/16x256x256.py b/configs/pixart/train/16x256x256.py new file mode 100644 index 00000000..b47731e2 --- /dev/null +++ b/configs/pixart/train/16x256x256.py @@ -0,0 +1,53 @@ +num_frames = 16 +frame_interval = 3 +image_size = (256, 256) + +# Define dataset +root = None +data_path = "CSV_PATH" +use_image_transform = False +num_workers = 4 + +# Define acceleration +dtype = "bf16" +grad_checkpoint = False +plugin = "zero2" +sp_size = 1 + +# Define model +model = dict( + type="PixArt-XL/2", + space_scale=0.5, + time_scale=1.0, + from_pretrained="PixArt-XL-2-512x512.pth", + enable_flashattn=True, + enable_layernorm_kernel=True, +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", +) +text_encoder = dict( + type="t5", + from_pretrained="./pretrained_models/t5_ckpts", + model_max_length=120, + shardformer=True, +) +scheduler = dict( + type="iddpm", + timestep_respacing="", +) + +# Others +seed = 42 +outputs = "outputs" +wandb = False + +epochs = 1000 +log_every = 10 +ckpt_every = 1000 +load = None + +batch_size = 8 +lr = 2e-5 +grad_clip = 1.0 diff --git a/configs/pixart/train/1x512x512.py b/configs/pixart/train/1x512x512.py new file mode 100644 index 00000000..619c9aaf --- /dev/null +++ b/configs/pixart/train/1x512x512.py @@ -0,0 +1,54 @@ +num_frames = 1 +frame_interval = 1 +image_size = (512, 512) + +# Define dataset +root = None +data_path = "CSV_PATH" +use_image_transform = True +num_workers = 4 + +# Define acceleration +dtype = "bf16" +grad_checkpoint = True +plugin = "zero2" +sp_size = 1 + +# Define model +model = dict( + type="PixArt-XL/2", + space_scale=1.0, + time_scale=1.0, + no_temporal_pos_emb=True, + from_pretrained="PixArt-XL-2-512x512.pth", + enable_flashattn=True, + enable_layernorm_kernel=True, +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", +) +text_encoder = dict( + type="t5", + from_pretrained="./pretrained_models/t5_ckpts", + model_max_length=120, + shardformer=True, +) +scheduler = dict( + type="iddpm", + timestep_respacing="", +) + +# Others +seed = 42 +outputs = "outputs" +wandb = False + +epochs = 1000 +log_every = 10 +ckpt_every = 1000 +load = None + +batch_size = 32 +lr = 2e-5 +grad_clip = 1.0 diff --git a/configs/pixart/train/64x512x512.py b/configs/pixart/train/64x512x512.py new file mode 100644 index 00000000..628cf254 --- /dev/null +++ b/configs/pixart/train/64x512x512.py @@ -0,0 +1,54 @@ +num_frames = 64 +frame_interval = 2 +image_size = (512, 512) + +# Define dataset +root = None +data_path = "CSV_PATH" +use_image_transform = False +num_workers = 4 + +# Define acceleration +dtype = "bf16" +grad_checkpoint = True +plugin = "zero2" +sp_size = 1 + +# Define model +model = dict( + type="PixArt-XL/2", + space_scale=1.0, + time_scale=2 / 3, + from_pretrained=None, + enable_flashattn=True, + enable_layernorm_kernel=True, +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", + micro_batch_size=128, +) +text_encoder = dict( + type="t5", + from_pretrained="./pretrained_models/t5_ckpts", + model_max_length=120, + shardformer=True, +) +scheduler = dict( + type="iddpm", + timestep_respacing="", +) + +# Others +seed = 42 +outputs = "outputs" +wandb = False + +epochs = 1000 +log_every = 10 +ckpt_every = 250 +load = None + +batch_size = 4 +lr = 2e-5 +grad_clip = 1.0 diff --git a/open_sora/__init__.py b/docs/README_zh.md similarity index 100% rename from open_sora/__init__.py rename to docs/README_zh.md diff --git a/docs/acceleration.md b/docs/acceleration.md new file mode 100644 index 00000000..3a0a68eb --- /dev/null +++ b/docs/acceleration.md @@ -0,0 +1,57 @@ +# Acceleration + +Open-Sora aims to provide a high-speed training framework for diffusion models. We can achieve **55%** training speed acceleration when training on **64 frames 512x512 videos**. Our framework support training **1min 1080p videos**. + +## Accelerated Transformer + +Open-Sora boosts the training speed by: + +- Kernal optimization including [flash attention](https://github.com/Dao-AILab/flash-attention), fused layernorm kernal, and the ones compiled by colossalAI. +- Hybrid parallelism including ZeRO. +- Gradient checkpointing for larger batch size. + +Our training speed on images is comparable to [OpenDiT](https://github.com/NUS-HPC-AI-Lab/OpenDiT), an project to accelerate DiT training. The training speed is measured on 8 H800 GPUs with batch size 128, image size 256x256. + +| Model | Throughput (img/s/GPU) | Throughput (tokens/s/GPU) | +| -------- | ---------------------- | ------------------------- | +| DiT | 100 | 26k | +| OpenDiT | 175 | 45k | +| OpenSora | 175 | 45k | + +## Efficient STDiT + +Our STDiT adopts spatial-temporal attention to model the video data. Compared with directly applying full attention on DiT, our STDiT is more efficient as the number of frames increases. Our current framework only supports sequence parallelism for very long sequence. + +The training speed is measured on 8 H800 GPUs with acceleration techniques applied, GC means gradient checkpointing. Both with T5 conditioning like PixArt. + +| Model | Setting | Throughput (sample/s/GPU) | Throughput (tokens/s/GPU) | +| ---------------- | -------------- | ------------------------- | ------------------------- | +| DiT | 16x256 (4k) | 7.20 | 29k | +| STDiT | 16x256 (4k) | 7.00 | 28k | +| DiT | 16x512 (16k) | 0.85 | 14k | +| STDiT | 16x512 (16k) | 1.45 | 23k | +| DiT (GC) | 64x512 (65k) | 0.08 | 5k | +| STDiT (GC) | 64x512 (65k) | 0.40 | 25k | +| STDiT (GC, sp=2) | 360x512 (370k) | 0.10 | 18k | + +With a 4x downsampling in the temporal dimension with Video-VAE, an 24fps video has 450 frames. The gap between the speed of STDiT (28k tokens/s) and DiT on images (up to 45k tokens/s) mainly comes from the T5 and VAE encoding, and temperal attention. + +## Accelerated Encoder (T5, VAE) + +During training, texts are encoded by T5, and videos are encoded by VAE. Typically there are two ways to accelerate the training: + +1. Preprocess text and video data in advance and save them to disk. +2. Encode text and video data during training, and accelerate the encoding process. + +For option 1, 120 tokens for one sample require 1M disk space, and a 64x64x64 latent requires 4M. Considering a training dataset with 10M video clips, the total disk space required is 50TB. Our storage system is not ready at this time for this scale of data. + +For option 2, we boost T5 speed and memory requirement. According to [OpenDiT](https://github.com/NUS-HPC-AI-Lab/OpenDiT), we find VAE consumes a large number of GPU memory. Thus we split batch size into smaller ones for VAE encoding. With both techniques, we can greatly accelerated the training speed. + +The training speed is measured on 8 H800 GPUs with STDiT. + +| Acceleration | Setting | Throughput (img/s/GPU) | Throughput (tokens/s/GPU) | +| ------------ | ------------- | ---------------------- | ------------------------- | +| Baseline | 16x256 (4k) | 6.16 | 25k | +| w. faster T5 | 16x256 (4k) | 7.00 | 29k | +| Baseline | 64x512 (65k) | 0.94 | 15k | +| w. both | 64x512 (65k) | 1.45 | 23k | diff --git a/docs/commands.md b/docs/commands.md new file mode 100644 index 00000000..28ee285d --- /dev/null +++ b/docs/commands.md @@ -0,0 +1,91 @@ +# Commands + +## Inference + +You can modify corresponding config files to change the inference settings. See more details [here](/docs/structure.md#inference-config-demos). + +### Inference with DiT pretrained on ImageNet + +The following command automatically downloads the pretrained weights on ImageNet and runs inference. + +```bash +python scripts/inference.py configs/dit/inference/1x256x256-class.py --ckpt-path DiT-XL-2-256x256.pt +``` + +### Inference with Latte pretrained on UCF101 + +The following command automatically downloads the pretrained weights on UCF101 and runs inference. + +```bash +python scripts/inference.py configs/latte/inference/16x256x256-class.py --ckpt-path Latte-XL-2-256x256-ucf101.pt +``` + +### Inference with PixArt-α pretrained weights + +Download T5 into `./pretrained_models` and run the following command. + +```bash +# 256x256 +torchrun --standalone --nproc_per_node 1 scripts/inference.py configs/pixart/inference/1x256x256.py --ckpt-path PixArt-XL-2-256x256.pth + +# 512x512 +torchrun --standalone --nproc_per_node 1 scripts/inference.py configs/pixart/inference/1x512x512.py --ckpt-path PixArt-XL-2-512x512.pth + +# 1024 multi-scale +torchrun --standalone --nproc_per_node 1 scripts/inference.py configs/pixart/inference/1x1024MS.py --ckpt-path PixArt-XL-2-1024MS.pth +``` + +### Inference with checkpoints saved during training + +During training, an experiment logging folder is created in `outputs` directory. Under each checpoint folder, e.g. `epoch12-global_step2000`, there is a `ema.pt` and the shared `model` folder. Run the following command to perform inference. + +```bash +# inference with ema model +torchrun --standalone --nproc_per_node 1 scripts/inference.py configs/opensora/inference/16x256x256.py --ckpt-path outputs/001-STDiT-XL-2/epoch12-global_step2000/ema.pt + +# inference with model +torchrun --standalone --nproc_per_node 1 scripts/inference.py configs/opensora/inference/16x256x256.py --ckpt-path outputs/001-STDiT-XL-2/epoch12-global_step2000 + +# inference with sequence parallelism +# sequence parallelism is enabled automatically when nproc_per_node is larger than 1 +torchrun --standalone --nproc_per_node 2 scripts/inference.py configs/opensora/inference/16x256x256.py --ckpt-path outputs/001-STDiT-XL-2/epoch12-global_step2000 +``` + +The second command will automatically generate a `model_ckpt.pt` file in the checkpoint folder. + +### Inference Hyperparameters + +1. DPM-solver is good at fast inference for images. However, the video result is not satisfactory. You can use it for fast demo purpose. + +```python +type="dmp-solver" +num_sampling_steps=20 +``` + +1. You can use [SVD](https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt)'s finetuned VAE decoder on videos for inference (consumes more memory). However, we do not see significant improvement in the video result. To use it, download [the pretrained weights](https://huggingface.co/maxin-cn/Latte/tree/main/t2v_required_models/vae_temporal_decoder) into `./pretrained_models/vae_temporal_decoder` and modify the config file as follows. + +```python +vae = dict( + type="VideoAutoencoderKLTemporalDecoder", + from_pretrained="pretrained_models/vae_temporal_decoder", +) + +## Training + +To resume training, run the following command. ``--load`` different from ``--ckpt-path`` as it loads the optimizer and dataloader states. + +```bash +torchrun --nnodes=1 --nproc_per_node=8 scripts/train.py configs/opensora/train/64x512x512.py --data-path YOUR_CSV_PATH --load YOUR_PRETRAINED_CKPT +``` + +To enable wandb logging, add `--wandb` to the command. + +```bash +WANDB_API_KEY=YOUR_WANDB_API_KEY torchrun --nnodes=1 --nproc_per_node=8 scripts/train.py configs/opensora/train/64x512x512.py --data-path YOUR_CSV_PATH --wandb True +``` + +You can modify corresponding config files to change the training settings. See more details [here](/docs/structure.md#training-config-demos). + +### Training Hyperparameters + +1. `dtype` is the data type for training. Only `fp16` and `bf16` are supported. ColossalAI automatically enables the mixed precision training for `fp16` and `bf16`. During training, we find `bf16` more stable. diff --git a/docs/datasets.md b/docs/datasets.md new file mode 100644 index 00000000..c06835b0 --- /dev/null +++ b/docs/datasets.md @@ -0,0 +1,28 @@ +# Datasets + +## Datasets used for now + +### HD-VG-130M + +[HD-VG-130M](https://github.com/daooshee/HD-VG-130M?tab=readme-ov-file) comprises 130M text-video pairs. The caption is generated by BLIP-2. We find the cut and the text quality are relatively poor. It contains 20 splits. For OpenSora 1.0, we use the first split. We plan to use the whole dataset and re-process it. + +### Inter4k + +[Inter4k](https://github.com/alexandrosstergiou/Inter4K) is a dataset containing 1k video clips with 4K resolution. The dataset is proposed for super-resolution tasks. We use the dataset for HQ training. The videos are processed as mentioned [here](/README.md#data-processing). + +### Pexels.com + +[Pexels.com](https://www.pexels.com/) is a website that provides free stock photos and videos. We collect 19K video clips from this website for HQ training. The videos are processed as mentioned [here](/README.md#data-processing). + +## Datasets watching list + +We are also watching the following datasets and considering using them in the future, which depends on our disk space and the quality of the dataset. + +| Name | Size | Description | +| ----------------- | ------------ | ----------------------------- | +| Panda-70M | 70M videos | High quality video-text pairs | +| WebVid-10M | 10M videos | Low quality | +| InternVid-10M-FLT | 10M videos | | +| EGO4D | 3670 hours | | +| OpenDV-YouTube | 1700 hours | | +| VidProM | 6.69M videos | | diff --git a/docs/report_v1.md b/docs/report_v1.md new file mode 100644 index 00000000..b3b8073c --- /dev/null +++ b/docs/report_v1.md @@ -0,0 +1,47 @@ +# Open-Sora v1 Report + +OpenAI's Sora is amazing at generating one minutes high quality videos. However, it reveals almost no information about its details. To make AI more "open", we are dedicated to build an open-source version of Sora. This report describes our first attempt to train a transformer-based video diffusion model. + +## Efficiency in choosing the architecture + +To lower the computational cost, we want to utilize existing VAE models. Sora uses spatial-temporal VAE to reduce the temporal dimensions. However, we found that there is no open-source high-quality spatial-temporal VAE model. [MAGVIT](https://github.com/google-research/magvit)'s 4x4x4 VAE is not open-sourced, while [VideoGPT](https://wilson1yan.github.io/videogpt/index.html)'s 2x4x4 VAE has a low quality in our experiments. Thus, we decided to use a 2D VAE (from [Stability-AI](https://huggingface.co/stabilityai/sd-vae-ft-mse-original)) in our first version. + +The video training involves a large amount of tokens. Considering 24fps 1min videos, we have 1440 frames. With VAE downsampling 4x and patch size downsampling 2x, we have 1440x1024≈1.5M tokens. Full attention on 1.5M tokens leads to a huge computational cost. Thus, we use spatial-temporal attention to reduce the cost following [Latte](https://github.com/Vchitect/Latte). + +As shown in the figure, we insert a temporal attention right after each spatial attention in STDiT (ST stands for spatial-temporal). This is similar to variant 3 in Latte's paper. However, we do not control a similar number of parameters for these variants. While Latte's paper claims their variant is better than variant 3, our experiments on 16x256x256 videos show that with same number of iterations, the performance ranks as: DiT (full) > STDiT (Sequential) > STDiT (Parallel) ≈ Latte. Thus, we choose STDiT (Sequential) out of efficiency. Speed benchmark is provided [here](/docs/acceleration.md#efficient-stdit). + +![Architecture Comparison](https://i0.imgs.ovh/2024/03/15/eLk9D.png) + +To focus on video generation, we hope to train the model based on a powerful image generation model. [PixArt-α](https://github.com/PixArt-alpha/PixArt-alpha) is an efficiently trained high-quality image generation model with T5-conditioned DiT structure. We initialize our model with PixArt-α and initialize the projection layer of inserted temporal attention with zero. This initialization preserves model's ability of image generation at beginning, while Latte's architecture cannot. The inserted attention increases the number of parameter from 580M to 724M. + +![Architecture](https://i0.imgs.ovh/2024/03/16/erC1d.png) + +Drawing from the success of PixArt-α and Stable Video Diffusion, we also adopt a progressive training strategy: 16x256x256 on 366K pretraining datasets, and then 16x256x256, 16x512x512, and 64x512x512 on 20K datasets. With scaled position embedding, this strategy greatly reduces the computational cost. + +We also try to use a 3D patch embedder in DiT. However, with 2x downsampling on temporal dimension, the generated videos have a low quality. Thus, we leave the downsampling to temporal VAE in our next version. For now, we sample at every 3 frames with 16 frames training and every 2 frames with 64 frames training. + +## Data is the key to high quality + +We find that the number and quality of data have a great impact on the quality of generated videos, even larger than the model architecture and training strategy. At this time, we only prepared the first split (366K video clips) from [HD-VG-130M](https://github.com/daooshee/HD-VG-130M). The quality of these videos varies greatly, and the captions are not that accurate. Thus, we further collect 20k relatively high quality videos from [Pexels](https://www.pexels.com/), which provides free license videos. We label the video with LLaVA, an image captioning model, with three frames and a designed prompt. With designed prompt, LLaVA can generate good quality of captions. + +![Caption](https://i0.imgs.ovh/2024/03/16/eXdvC.png) + +As we lay more emphasis on the quality of data, we prepare to collect more data and build a video preprocessing pipeline in our next version. + +## Training Details + +With a limited training budgets, we made only a few exploration. We find learning rate 1e-4 is too large and scales down to 2e-5. When training with a large batch size, we find `fp16` less stable than `bf16` and may lead to generation failure. Thus, we switch to `bf16` for training on 64x512x512. For other hyper-parameters, we follow previous works. + +## Loss curves + +16x256x256 Pretraining Loss Curve + +![16x256x256 Pretraining Loss Curve](https://i0.imgs.ovh/2024/03/16/erXQj.png) + +16x256x256 HQ Training Loss Curve + +![16x256x256 HQ Training Loss Curve](https://i0.imgs.ovh/2024/03/16/ernXv.png) + +16x512x512 HQ Training Loss Curve + +![16x512x512 HQ Training Loss Curve](https://i0.imgs.ovh/2024/03/16/erHBe.png) diff --git a/docs/structure.md b/docs/structure.md new file mode 100644 index 00000000..0fc087ee --- /dev/null +++ b/docs/structure.md @@ -0,0 +1,178 @@ +# Repo & Config Structure + +## Repo Structure + +```plaintext +Open-Sora +├── README.md +├── docs +│ ├── acceleration.md -> Acceleration & Speed benchmark +│ ├── command.md -> Commands for training & inference +│ ├── datasets.md -> Datasets used in this project +│ ├── structure.md -> This file +│ └── report_v1.md -> Report for Open-Sora v1 +├── scripts +│ ├── train.py -> diffusion training script +│ └── inference.py -> Report for Open-Sora v1 +├── configs -> Configs for training & inference +├── opensora +│ ├── __init__.py +│ ├── registry.py -> Registry helper +│   ├── acceleration -> Acceleration related code +│   ├── dataset -> Dataset related code +│   ├── models +│   │   ├── layers -> Common layers +│   │   ├── vae -> VAE as image encoder +│   │   ├── text_encoder -> Text encoder +│   │   │   ├── classes.py -> Class id encoder (inference only) +│   │   │   ├── clip.py -> CLIP encoder +│   │   │   └── t5.py -> T5 encoder +│   │   ├── dit +│   │   ├── latte +│   │   ├── pixart +│   │   └── stdit -> Our STDiT related code +│   ├── schedulers -> Diffusion shedulers +│   │   ├── iddpm -> IDDPM for training and inference +│   │ └── dpms -> DPM-Solver for fast inference +│ └── utils +└── tools -> Tools for data processing and more +``` + +## Configs + +Our config files follows [MMEgine](https://github.com/open-mmlab/mmengine). MMEngine will reads the config file (a `.py` file) and parse it into a dictionary-like object. + +```plaintext +Open-Sora +└── configs -> Configs for training & inference + ├── opensora -> STDiT related configs + │ ├── inference + │ │ ├── 16x256x256.py -> Sample videos 16 frames 256x256 + │ │ ├── 16x512x512.py -> Sample videos 16 frames 512x512 + │ │ └── 64x512x512.py -> Sample videos 64 frames 512x512 + │ └── train + │ ├── 16x256x256.py -> Train on videos 16 frames 256x256 + │ ├── 16x256x256.py -> Train on videos 16 frames 256x256 + │ └── 64x512x512.py -> Train on videos 64 frames 512x512 + ├── dit -> DiT related configs +    │   ├── inference +    │   │   ├── 1x256x256-class.py -> Sample images with ckpts from DiT +    │   │   ├── 1x256x256.py -> Sample images with clip condition +    │   │   └── 16x256x256.py -> Sample videos +    │   └── train +    │     ├── 1x256x256.py -> Train on images with clip condition +    │      └── 16x256x256.py -> Train on videos + ├── latte -> Latte related configs + └── pixart -> PixArt related configs +``` + +## Inference config demos + +To change the inference settings, you can directly modify the corresponding config file. Or you can pass arguments to overwrite the config file ([config_utils.py](/opensora/utils/config_utils.py)). To change sampling prompts, you should modify the `.txt` file passed to the `--prompt_path` argument. + +```plaintext +--prompt_path ./assets/texts/t2v_samples.txt -> prompt_path +--ckpt-path ./path/to/your/ckpt.pth -> model["from_pretrained"] +``` + +The explanation of each field is provided below. + +```python +# Define sampling size +num_frames = 64 # number of frames +fps = 24 // 2 # frames per second (divided by 2 for frame_interval=2) +image_size = (512, 512) # image size (height, width) + +# Define model +model = dict( + type="STDiT-XL/2", # Select model type (STDiT-XL/2, DiT-XL/2, etc.) + space_scale=1.0, # (Optional) Space positional encoding scale (new height / old height) + time_scale=2 / 3, # (Optional) Time positional encoding scale (new frame_interval / old frame_interval) + enable_flashattn=True, # (Optional) Speed up training and inference with flash attention + enable_layernorm_kernel=True, # (Optional) Speed up training and inference with fused kernel + from_pretrained="PRETRAINED_MODEL", # (Optional) Load from pretrained model + no_temporal_pos_emb=True, # (Optional) Disable temporal positional encoding (for image) +) +vae = dict( + type="VideoAutoencoderKL", # Select VAE type + from_pretrained="stabilityai/sd-vae-ft-ema", # Load from pretrained VAE + micro_batch_size=128, # VAE with micro batch size to save memory +) +text_encoder = dict( + type="t5", # Select text encoder type (t5, clip) + from_pretrained="./pretrained_models/t5_ckpts", # Load from pretrained text encoder + model_max_length=120, # Maximum length of input text +) +scheduler = dict( + type="iddpm", # Select scheduler type (iddpm, dpm-solver) + num_sampling_steps=100, # Number of sampling steps + cfg_scale=7.0, # hyper-parameter for classifier-free diffusion +) +dtype = "fp16" # Computation type (fp16, fp32, bf16) + +# Other settings +batch_size = 1 # batch size +seed = 42 # random seed +prompt_path = "./assets/texts/t2v_samples.txt" # path to prompt file +save_dir = "./samples" # path to save samples +``` + +## Training config demos + +```python +# Define sampling size +num_frames = 64 +frame_interval = 2 # sample every 2 frames +image_size = (512, 512) + +# Define dataset +root = None # root path to the dataset +data_path = "CSV_PATH" # path to the csv file +use_image_transform = False # True if training on images +num_workers = 4 # number of workers for dataloader + +# Define acceleration +dtype = "bf16" # Computation type (fp16, bf16) +grad_checkpoint = True # Use gradient checkpointing +plugin = "zero2" # Plugin for distributed training (zero2, zero2-seq) +sp_size = 1 # Sequence parallelism size (1 for no sequence parallelism) + +# Define model +model = dict( + type="STDiT-XL/2", + space_scale=1.0, + time_scale=2 / 3, + from_pretrained="YOUR_PRETRAINED_MODEL", + enable_flashattn=True, # Enable flash attention + enable_layernorm_kernel=True, # Enable layernorm kernel +) +vae = dict( + type="VideoAutoencoderKL", + from_pretrained="stabilityai/sd-vae-ft-ema", + micro_batch_size=128, +) +text_encoder = dict( + type="t5", + from_pretrained="./pretrained_models/t5_ckpts", + model_max_length=120, + shardformer=True, # Enable shardformer for T5 acceleration +) +scheduler = dict( + type="iddpm", + timestep_respacing="", # Default 1000 timesteps +) + +# Others +seed = 42 +outputs = "outputs" # path to save checkpoints +wandb = False # Use wandb for logging + +epochs = 1000 # number of epochs (just large enough, kill when satisfied) +log_every = 10 +ckpt_every = 250 +load = None # path to resume training + +batch_size = 4 +lr = 2e-5 +grad_clip = 1.0 # gradient clipping +``` diff --git a/download.py b/download.py deleted file mode 100644 index f4c37007..00000000 --- a/download.py +++ /dev/null @@ -1,50 +0,0 @@ -# Copyright (c) Meta Platforms, Inc. and affiliates. -# All rights reserved. - -# This source code is licensed under the license found in the -# LICENSE file in the root directory of this source tree. - -""" -Functions for downloading pre-trained DiT models -""" -import os - -import torch -from torchvision.datasets.utils import download_url - -pretrained_models = {"DiT-XL-2-512x512.pt", "DiT-XL-2-256x256.pt"} - - -def find_model(model_name): - """ - Finds a pre-trained DiT model, downloading it if necessary. Alternatively, loads a model from a local path. - """ - if model_name in pretrained_models: # Find/download our pre-trained DiT checkpoints - return download_model(model_name) - else: # Load a custom DiT checkpoint: - assert os.path.isfile(model_name), f"Could not find DiT checkpoint at {model_name}" - checkpoint = torch.load(model_name, map_location=lambda storage, loc: storage) - if "ema" in checkpoint: # supports checkpoints from train.py - checkpoint = checkpoint["ema"] - return checkpoint - - -def download_model(model_name): - """ - Downloads a pre-trained DiT model from the web. - """ - assert model_name in pretrained_models - local_path = f"pretrained_models/{model_name}" - if not os.path.isfile(local_path): - os.makedirs("pretrained_models", exist_ok=True) - web_path = f"https://dl.fbaipublicfiles.com/DiT/models/{model_name}" - download_url(web_path, "pretrained_models") - model = torch.load(local_path, map_location=lambda storage, loc: storage) - return model - - -if __name__ == "__main__": - # Download all DiT checkpoints - for model in pretrained_models: - download_model(model) - print("Done.") diff --git a/open_sora/diffusion/__init__.py b/open_sora/diffusion/__init__.py deleted file mode 100644 index 427068de..00000000 --- a/open_sora/diffusion/__init__.py +++ /dev/null @@ -1,40 +0,0 @@ -# Modified from OpenAI's diffusion repos -# GLIDE: https://github.com/openai/glide-text2im/blob/main/glide_text2im/gaussian_diffusion.py -# ADM: https://github.com/openai/guided-diffusion/blob/main/guided_diffusion -# IDDPM: https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py - -from . import gaussian_diffusion as gd -from .respace import SpacedDiffusion, space_timesteps - - -def create_diffusion( - timestep_respacing, - noise_schedule="linear", - use_kl=False, - sigma_small=False, - predict_xstart=False, - learn_sigma=True, - rescale_learned_sigmas=False, - diffusion_steps=1000, -): - betas = gd.get_named_beta_schedule(noise_schedule, diffusion_steps) - if use_kl: - loss_type = gd.LossType.RESCALED_KL - elif rescale_learned_sigmas: - loss_type = gd.LossType.RESCALED_MSE - else: - loss_type = gd.LossType.MSE - if timestep_respacing is None or timestep_respacing == "": - timestep_respacing = [diffusion_steps] - return SpacedDiffusion( - use_timesteps=space_timesteps(diffusion_steps, timestep_respacing), - betas=betas, - model_mean_type=(gd.ModelMeanType.EPSILON if not predict_xstart else gd.ModelMeanType.START_X), - model_var_type=( - (gd.ModelVarType.FIXED_LARGE if not sigma_small else gd.ModelVarType.FIXED_SMALL) - if not learn_sigma - else gd.ModelVarType.LEARNED_RANGE - ), - loss_type=loss_type - # rescale_timesteps=rescale_timesteps, - ) diff --git a/open_sora/modeling/__init__.py b/open_sora/modeling/__init__.py deleted file mode 100644 index f1355ce4..00000000 --- a/open_sora/modeling/__init__.py +++ /dev/null @@ -1,4 +0,0 @@ -from .dit import DiT, DiT_models -from .latte import LatteT2V - -__all__ = ["DiT_models", "DiT", "LatteT2V"] diff --git a/open_sora/modeling/dit/__init__.py b/open_sora/modeling/dit/__init__.py deleted file mode 100644 index 73813b5d..00000000 --- a/open_sora/modeling/dit/__init__.py +++ /dev/null @@ -1,3 +0,0 @@ -from .dit import SUPPORTED_MODEL_ARCH, SUPPORTED_SEQ_PARALLEL_MODES, DiT, DiT_models - -__all__ = ["DiT_models", "DiT", "SUPPORTED_SEQ_PARALLEL_MODES", "SUPPORTED_MODEL_ARCH"] diff --git a/open_sora/modeling/dit/attn.py b/open_sora/modeling/dit/attn.py deleted file mode 100644 index 6ff95f25..00000000 --- a/open_sora/modeling/dit/attn.py +++ /dev/null @@ -1,276 +0,0 @@ -from typing import Optional - -import torch -import torch.distributed as dist -import torch.nn as nn -import torch.nn.functional as F -from colossalai.logging import get_dist_logger -from colossalai.shardformer.layer._operation import gather_forward_split_backward - -from open_sora.utils.comm import all_to_all, async_all_gather_proj_for_two - - -class CrossAttention(nn.Module): - r""" - A cross attention layer. - - Parameters: - query_dim (`int`): The number of channels in the query. - cross_attention_dim (`int`, *optional*): - The number of channels in the context. If not given, defaults to `query_dim`. - num_heads (`int`, *optional*, defaults to 8): The number of heads to use for multi-head attention. - head_dim (`int`, *optional*, defaults to 64): The number of channels in each head. - dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. - bias (`bool`, *optional*, defaults to False): - Set to `True` for the query, key, and value linear layers to contain a bias parameter. - """ - - def __init__( - self, - query_dim: int, - cross_attention_dim: Optional[int] = None, - num_heads: int = 8, - head_dim: int = 64, - dropout: float = 0.0, - bias=False, - sdpa=True, - ): - super().__init__() - self.hidden_size = head_dim * num_heads - cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim - - self.scale = head_dim**-0.5 - self.num_heads = num_heads - self.head_dim = head_dim - self.sdpa = sdpa - - self.to_q = nn.Linear(query_dim, self.hidden_size, bias=bias) - self.to_k = nn.Linear(cross_attention_dim, self.hidden_size, bias=bias) - self.to_v = nn.Linear(cross_attention_dim, self.hidden_size, bias=bias) - - self.to_out = nn.Sequential(nn.Linear(self.hidden_size, query_dim), nn.Dropout(dropout)) - - def forward(self, hidden_states, context=None, mask=None): - bsz, q_len, _ = hidden_states.shape - - query = self.to_q(hidden_states) - context = context if context is not None else hidden_states - kv_seq_len = context.shape[1] - key = self.to_k(context) - value = self.to_v(context) - - # [B, S, H, D] - query = query.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) - key = key.view(bsz, kv_seq_len, self.num_heads, self.head_dim).transpose(1, 2) - value = value.view(bsz, kv_seq_len, self.num_heads, self.head_dim).transpose(1, 2) - - if mask is not None: - assert mask.shape == (bsz, 1, q_len, kv_seq_len) - if self.sdpa: - attn_output = F.scaled_dot_product_attention(query, key, value, attn_mask=mask, scale=self.scale) - else: - attn_weights = torch.matmul(query, key.transpose(2, 3)) / self.scale - assert attn_weights.shape == (bsz, self.num_heads, q_len, kv_seq_len) - if mask is not None: - attn_weights = attn_weights + mask - attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype) - attn_output = torch.matmul(attn_weights, value) - assert attn_output.shape == (bsz, self.num_heads, q_len, self.head_dim) - attn_output = attn_output.transpose(1, 2).contiguous() - attn_output = attn_output.reshape(bsz, q_len, self.hidden_size) - attn_output = self.to_out(attn_output) - return attn_output - - -class SeqParallelCrossAttention(CrossAttention): - r""" - A cross attention layer. - - Parameters: - query_dim (`int`): The number of channels in the query. - cross_attention_dim (`int`, *optional*): - The number of channels in the context. If not given, defaults to `query_dim`. - num_heads (`int`, *optional*, defaults to 8): The number of heads to use for multi-head attention. - head_dim (`int`, *optional*, defaults to 64): The number of channels in each head. - dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. - bias (`bool`, *optional*, defaults to False): - Set to `True` for the query, key, and value linear layers to contain a bias parameter. - """ - - def __init__( - self, - query_dim: int, - cross_attention_dim: Optional[int] = None, - num_heads: int = 8, - head_dim: int = 64, - dropout: float = 0.0, - bias=False, - sdpa=True, - seq_parallel_group=None, - ): - super().__init__( - query_dim, - cross_attention_dim, - num_heads, - head_dim, - dropout, - bias, - sdpa, - ) - self.seq_parallel_group = seq_parallel_group - self.seq_parallel_size = dist.get_world_size(self.seq_parallel_group) if seq_parallel_group is not None else 1 - assert self.num_heads % self.seq_parallel_size == 0 - - def forward(self, hidden_states, context=None, mask=None): - bsz, q_len, _ = hidden_states.shape - - query = self.to_q(hidden_states) - context = context if context is not None else hidden_states - kv_seq_len = context.shape[1] - key = self.to_k(context) - value = self.to_v(context) - - # [B, S/P, H] -> [B, S, H/P] - num_heads_parallel = self.num_heads // self.seq_parallel_size - hidden_size_parallel = self.hidden_size // self.seq_parallel_size - if self.seq_parallel_size > 1: - query = all_to_all(query, self.seq_parallel_group, scatter_dim=2, gather_dim=1) - key = all_to_all(key, self.seq_parallel_group, scatter_dim=2, gather_dim=1) - value = all_to_all(value, self.seq_parallel_group, scatter_dim=2, gather_dim=1) - - q_len *= self.seq_parallel_size - kv_seq_len *= self.seq_parallel_size - - # [B, S, H/P] -> [B, S, N/P, D] -> [B, N/P, S, D] - query = query.view(bsz, q_len, num_heads_parallel, self.head_dim).transpose(1, 2) - key = key.view(bsz, kv_seq_len, num_heads_parallel, self.head_dim).transpose(1, 2) - value = value.view(bsz, kv_seq_len, num_heads_parallel, self.head_dim).transpose(1, 2) - - if mask is not None: - assert mask.shape == (bsz, 1, q_len, kv_seq_len) - if self.sdpa: - attn_output = F.scaled_dot_product_attention(query, key, value, attn_mask=mask, scale=self.scale) - else: - attn_weights = torch.matmul(query, key.transpose(2, 3)) / self.scale - assert attn_weights.shape == (bsz, num_heads_parallel, q_len, kv_seq_len) - if mask is not None: - attn_weights = attn_weights + mask - attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype) - attn_output = torch.matmul(attn_weights, value) - assert attn_output.shape == (bsz, num_heads_parallel, q_len, self.head_dim) - attn_output = attn_output.transpose(1, 2).contiguous() - attn_output = attn_output.reshape(bsz, q_len, hidden_size_parallel) - # [B, S, H/P] -> [B, S/P, H] - if self.seq_parallel_size > 1: - attn_output = all_to_all(attn_output, self.seq_parallel_group, scatter_dim=1, gather_dim=2) - attn_output = self.to_out(attn_output) - return attn_output - - -class FastSeqParallelCrossAttention(SeqParallelCrossAttention): - def __init__( - self, - query_dim: int, - cross_attention_dim: Optional[int] = None, - num_heads: int = 8, - head_dim: int = 64, - dropout: float = 0, - bias=False, - sdpa=True, - seq_parallel_group=None, - overlap=False, - ): - super().__init__( - query_dim, - cross_attention_dim, - num_heads, - head_dim, - dropout, - bias, - sdpa, - seq_parallel_group, - ) - self.seq_parallel_rank = dist.get_rank(self.seq_parallel_group) if seq_parallel_group is not None else 0 - self.sequence_parallel_param_slice = slice( - self.hidden_size // self.seq_parallel_size * self.seq_parallel_rank, - self.hidden_size // self.seq_parallel_size * (self.seq_parallel_rank + 1), - ) - if overlap and self.seq_parallel_size != 2: - logger = get_dist_logger() - logger.warning( - "FastSeqParallelCrossAttention only supports overlap with seq_parallel_size=2. Fallback to non-overlap", - ranks=[0], - ) - overlap = False - self.overlap = overlap - - def _get_sliced_params(self, proj_layer: nn.Linear): - bias = bias = proj_layer.bias[self.sequence_parallel_param_slice] if proj_layer.bias is not None else None - return proj_layer.weight[self.sequence_parallel_param_slice], bias - - def _proj(self, x: torch.Tensor, proj_layer: nn.Linear): - return F.linear( - x, - *self._get_sliced_params(proj_layer), - ) - - def forward(self, hidden_states, context=None, mask=None): - bsz, q_len, _ = hidden_states.shape - - context = context if context is not None else hidden_states - kv_seq_len = context.shape[1] - if self.seq_parallel_size > 1: - if self.overlap and self.seq_parallel_size == 2: - query, key, value = async_all_gather_proj_for_two( - hidden_states, - context, - *self._get_sliced_params(self.to_q), - *self._get_sliced_params(self.to_k), - *self._get_sliced_params(self.to_v), - dim=1, - process_group=self.seq_parallel_group, - sp_size=self.seq_parallel_size, - sp_rank=self.seq_parallel_rank, - ) - else: - # [B, S/P, H] -> [B, S, H] - hidden_states = gather_forward_split_backward(hidden_states, 1, self.seq_parallel_group) - context = gather_forward_split_backward(context, 1, self.seq_parallel_group) - query = self._proj(hidden_states, self.to_q) - key = self._proj(context, self.to_k) - value = self._proj(context, self.to_v) - else: - query = self.to_q(hidden_states) - key = self.to_k(context) - value = self.to_v(context) - # output is [B, S, H/P] - - num_heads_parallel = self.num_heads // self.seq_parallel_size - hidden_size_parallel = self.hidden_size // self.seq_parallel_size - q_len *= self.seq_parallel_size - kv_seq_len *= self.seq_parallel_size - - # [B, S, H/P] -> [B, S, N/P, D] -> [B, N/P, S, D] - query = query.view(bsz, q_len, num_heads_parallel, self.head_dim).transpose(1, 2) - key = key.view(bsz, kv_seq_len, num_heads_parallel, self.head_dim).transpose(1, 2) - value = value.view(bsz, kv_seq_len, num_heads_parallel, self.head_dim).transpose(1, 2) - - if mask is not None: - assert mask.shape == (bsz, 1, q_len, kv_seq_len) - if self.sdpa: - attn_output = F.scaled_dot_product_attention(query, key, value, attn_mask=mask, scale=self.scale) - else: - attn_weights = torch.matmul(query, key.transpose(2, 3)) / self.scale - assert attn_weights.shape == (bsz, num_heads_parallel, q_len, kv_seq_len) - if mask is not None: - attn_weights = attn_weights + mask - attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype) - attn_output = torch.matmul(attn_weights, value) - assert attn_output.shape == (bsz, num_heads_parallel, q_len, self.head_dim) - attn_output = attn_output.transpose(1, 2).contiguous() - attn_output = attn_output.reshape(bsz, q_len, hidden_size_parallel) - # [B, S, H/P] -> [B, S/P, H] - if self.seq_parallel_size > 1: - attn_output = all_to_all(attn_output, self.seq_parallel_group, scatter_dim=1, gather_dim=2) - attn_output = self.to_out(attn_output) - return attn_output diff --git a/open_sora/modeling/dit/dit.py b/open_sora/modeling/dit/dit.py deleted file mode 100644 index 6b7bd00c..00000000 --- a/open_sora/modeling/dit/dit.py +++ /dev/null @@ -1,659 +0,0 @@ -# Copyright (c) Meta Platforms, Inc. and affiliates. -# All rights reserved. - -# This source code is licensed under the license found in the -# LICENSE file in the root directory of this source tree. -# -------------------------------------------------------- -# References: -# GLIDE: https://github.com/openai/glide-text2im -# MAE: https://github.com/facebookresearch/mae/blob/main/models_mae.py -# -------------------------------------------------------- - -import math -from typing import Callable, Optional - -import numpy as np -import torch -import torch.distributed as dist -import torch.nn as nn -import torch.nn.functional as F -from timm.models.vision_transformer import Mlp - -from open_sora.utils.comm import gather_seq, split_seq - -from .attn import ( - CrossAttention, - FastSeqParallelCrossAttention, - SeqParallelCrossAttention, -) - -SUPPORTED_SEQ_PARALLEL_MODES = ["ulysses", "fastseq"] -SUPPORTED_MODEL_ARCH = ["adaln", "cross-attn", "token-concat"] - - -def modulate(x, shift, scale): - return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1) - - -################################################################################# -# Embedding Layers for Timesteps and Class Labels # -################################################################################# - - -class TimestepEmbedder(nn.Module): - """ - Embeds scalar timesteps into vector representations. - """ - - def __init__(self, hidden_size, frequency_embedding_size=256): - super().__init__() - self.mlp = nn.Sequential( - nn.Linear(frequency_embedding_size, hidden_size, bias=True), - nn.SiLU(), - nn.Linear(hidden_size, hidden_size, bias=True), - ) - self.frequency_embedding_size = frequency_embedding_size - - @staticmethod - def timestep_embedding(t, dim, max_period=10000): - """ - Create sinusoidal timestep embeddings. - :param t: a 1-D Tensor of N indices, one per batch element. - These may be fractional. - :param dim: the dimension of the output. - :param max_period: controls the minimum frequency of the embeddings. - :return: an (N, D) Tensor of positional embeddings. - """ - # https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py - half = dim // 2 - freqs = torch.exp( - -math.log(max_period) - * torch.arange(start=0, end=half, dtype=torch.float32) - / half - ).to(device=t.device) - args = t[:, None].float() * freqs[None] - embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1) - if dim % 2: - embedding = torch.cat( - [embedding, torch.zeros_like(embedding[:, :1])], dim=-1 - ) - return embedding - - def forward(self, t): - t_freq = self.timestep_embedding(t, self.frequency_embedding_size).to( - self.mlp[0].weight.dtype - ) - t_emb = self.mlp(t_freq) - return t_emb - - -class LabelEmbedder(nn.Module): - """ - Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance. - """ - - def __init__(self, num_classes, hidden_size, dropout_prob): - super().__init__() - use_cfg_embedding = dropout_prob > 0 - self.embedding_table = nn.Embedding( - num_classes + use_cfg_embedding, hidden_size - ) - self.num_classes = num_classes - self.dropout_prob = dropout_prob - - def token_drop(self, labels, force_drop_ids=None): - """ - Drops labels to enable classifier-free guidance. - """ - if force_drop_ids is None: - drop_ids = ( - torch.rand(labels.shape[0], device=labels.device) < self.dropout_prob - ) - else: - drop_ids = force_drop_ids == 1 - labels = torch.where(drop_ids, self.num_classes, labels) - return labels - - def forward(self, labels, train, force_drop_ids=None): - use_dropout = self.dropout_prob > 0 - if (train and use_dropout) or (force_drop_ids is not None): - labels = self.token_drop(labels, force_drop_ids) - embeddings = self.embedding_table(labels) - return embeddings - - -class PatchEmbedder(nn.Module): - """Patch Embedding Layer for flat 4D video tensors.""" - - def __init__( - self, - patch_size: int = 16, - in_chans: int = 3, - embed_dim: int = 768, - norm_layer: Optional[Callable] = None, - bias: bool = True, - ) -> None: - super().__init__() - self.patch_size = patch_size - self.proj = nn.Conv2d( - in_chans, embed_dim, kernel_size=patch_size, stride=patch_size, bias=bias - ) - self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity() - - def forward(self, x: torch.Tensor) -> torch.Tensor: - # [B, S, C, P, P] -> [B, S, C*P*P] - x = x.view(*x.shape[:2], -1) - out = F.linear( - x, self.proj.weight.view(self.proj.weight.shape[0], -1), self.proj.bias - ) - out = self.norm(out) - # [B, S, H] - return out - - -class TextEmbedder(nn.Module): - def __init__( - self, - in_features: int, - embed_dim: int = 768, - bias: bool = True, - dropout_prob: float = 0.0, - use_proj: bool = True, - ) -> None: - super().__init__() - self.dropout_prob = dropout_prob - self.use_proj = use_proj - if self.use_proj: - self.proj = nn.Linear(in_features, embed_dim, bias=bias) - - def drop_sample(self, x: torch.Tensor) -> torch.Tensor: - drop_ids = ( - torch.rand(x.shape[0], *([1] * (x.ndim - 1)), device=x.device) - < self.dropout_prob - ) - x = torch.where(drop_ids, torch.zeros_like(x), x) - return x - - def forward(self, x: torch.Tensor) -> torch.Tensor: - use_dropout = self.dropout_prob > 0 - if self.training and use_dropout: - x = self.drop_sample(x) - if self.use_proj: - # [B, S, C] -> [B, S, H] - x = self.proj(x) - return x - - -class PositionEmbedding(nn.Module): - def __init__(self, dim: int, max_position_embeddings=262114) -> None: - super().__init__() - self.dim = dim - self.max_position_embeddings = max_position_embeddings - self._set_pos_embed_cache(max_position_embeddings) - - def _set_pos_embed_cache(self, seq_len: int, device="cpu", dtype=torch.float): - self.max_seq_len_cached = seq_len - pos_embed = get_2d_sincos_pos_embed(self.dim, math.ceil(seq_len**0.5)) - pos_embed = torch.from_numpy(pos_embed).to(device=device, dtype=dtype) - # [S, H] - self.register_buffer("pos_embed_cache", pos_embed, persistent=False) - - def forward(self, x: torch.Tensor) -> torch.Tensor: - # [B, S, H] - seq_len = x.shape[1] - if seq_len > self.max_seq_len_cached: - self._set_pos_embed_cache(seq_len, x.device, x.dtype) - pos_embed = self.pos_embed_cache[None, :seq_len] - return pos_embed - - -################################################################################# -# Core DiT Model # -################################################################################# - - -class DiTBlock(nn.Module): - """ - A DiT block with adaptive layer norm zero (adaLN-Zero) conditioning. - """ - - def __init__( - self, - hidden_size, - num_heads, - cross_attention_dim=None, - mlp_ratio=4.0, - seq_parallel_group=None, - seq_parallel_mode="ulysses", - seq_parallel_overlap=False, - ): - super().__init__() - self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) - attn_kwargs = {} - if seq_parallel_group is not None: - attn_kwargs["seq_parallel_group"] = seq_parallel_group - if seq_parallel_mode == "ulysses": - attn_cls = SeqParallelCrossAttention - elif seq_parallel_mode == "fastseq": - attn_cls = FastSeqParallelCrossAttention - attn_kwargs["overlap"] = seq_parallel_overlap - else: - raise ValueError( - f"seq_parallel_mode must be one of {SUPPORTED_SEQ_PARALLEL_MODES}" - ) - else: - attn_cls = CrossAttention - self.attn = attn_cls( - query_dim=hidden_size, - cross_attention_dim=cross_attention_dim, - num_heads=num_heads, - head_dim=hidden_size // num_heads, - bias=True, - sdpa=True, - **attn_kwargs, - ) - - self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) - mlp_hidden_dim = int(hidden_size * mlp_ratio) - approx_gelu = lambda: nn.GELU(approximate="tanh") - self.mlp = Mlp( - in_features=hidden_size, - hidden_features=mlp_hidden_dim, - act_layer=approx_gelu, - drop=0, - ) - self.mlp = Mlp( - in_features=hidden_size, - hidden_features=mlp_hidden_dim, - act_layer=approx_gelu, - drop=0, - ) - self.adaLN_modulation = nn.Sequential( - nn.SiLU(), nn.Linear(hidden_size, 6 * hidden_size, bias=True) - ) - - def forward(self, x, attention_mask, t, context=None): - shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ( - self.adaLN_modulation(t).chunk(6, dim=1) - ) - x = x + gate_msa.unsqueeze(1) * self.attn( - modulate(self.norm1(x), shift_msa, scale_msa), context, attention_mask - ) - x = x + gate_mlp.unsqueeze(1) * self.mlp( - modulate(self.norm2(x), shift_mlp, scale_mlp) - ) - return x - - -class FinalLayer(nn.Module): - """ - The final layer of DiT. - """ - - def __init__(self, hidden_size, patch_size, out_channels): - super().__init__() - self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) - self.linear = nn.Linear( - hidden_size, patch_size * patch_size * out_channels, bias=True - ) - self.patch_size = patch_size - self.adaLN_modulation = nn.Sequential( - nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True) - ) - - def unpatchify(self, x): - b, s, h = x.shape - return x.view(b, s, -1, self.patch_size, self.patch_size) - - def forward(self, x, t): - shift, scale = self.adaLN_modulation(t).chunk(2, dim=1) - x = modulate(self.norm_final(x), shift, scale) - x = self.linear(x) - x = self.unpatchify(x) - return x - - -class DiT(nn.Module): - """ - Diffusion model with a Transformer backbone. - """ - - def __init__( - self, - patch_size=2, - in_channels=3, - text_embed_dim=512, - hidden_size=1152, - depth=28, - num_heads=16, - mlp_ratio=4.0, - max_num_embeddings=256 * 1024, - text_dropout_prob=0.1, - learn_sigma=True, - seq_parallel_group=None, - seq_parallel_mode="ulysses", - seq_parallel_overlap=False, - model_arch="cross-attn", - ): - super().__init__() - assert model_arch in SUPPORTED_MODEL_ARCH - self.model_arch = model_arch - self.grad_checkpointing = False - self.learn_sigma = learn_sigma - self.in_channels = in_channels - self.out_channels = in_channels * 2 if learn_sigma else in_channels - self.patch_size = patch_size - self.num_heads = num_heads - self.seq_parallel_group = seq_parallel_group - self.seq_parallel_size = ( - dist.get_world_size(self.seq_parallel_group) - if seq_parallel_group is not None - else 1 - ) - self.seq_parallel_rank = ( - dist.get_rank(self.seq_parallel_group) - if seq_parallel_group is not None - else 0 - ) - - self.video_embedder = PatchEmbedder( - patch_size, in_channels, hidden_size, bias=True - ) - self.t_embedder = TimestepEmbedder(hidden_size) - self.pos_embed = PositionEmbedding(hidden_size, max_num_embeddings) - self.text_embedder = TextEmbedder( - text_embed_dim, - hidden_size, - bias=True, - dropout_prob=text_dropout_prob, - use_proj=model_arch != "cross-attn", - ) - if model_arch != "cross-attn": - cross_attn_dim = None - else: - cross_attn_dim = text_embed_dim - - self.blocks = nn.ModuleList( - [ - DiTBlock( - hidden_size, - num_heads, - cross_attn_dim, - mlp_ratio=mlp_ratio, - seq_parallel_group=seq_parallel_group, - seq_parallel_mode=seq_parallel_mode, - seq_parallel_overlap=seq_parallel_overlap, - ) - for _ in range(depth) - ] - ) - self.final_layer = FinalLayer(hidden_size, patch_size, self.out_channels) - self.initialize_weights() - - def initialize_weights(self): - # Initialize transformer layers: - def _basic_init(module): - if isinstance(module, nn.Linear): - torch.nn.init.xavier_uniform_(module.weight) - if module.bias is not None: - nn.init.constant_(module.bias, 0) - - self.apply(_basic_init) - - # Initialize text embedding layer - if self.text_embedder.use_proj: - nn.init.normal_(self.text_embedder.proj.weight, std=0.02) - - # Initialize timestep embedding MLP: - nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02) - nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02) - - # Zero-out adaLN modulation layers - for block in self.blocks: - nn.init.constant_(block.adaLN_modulation[-1].weight, 0) - nn.init.constant_(block.adaLN_modulation[-1].bias, 0) - - # Zero-out output - nn.init.constant_(self.final_layer.adaLN_modulation[-1].weight, 0) - nn.init.constant_(self.final_layer.adaLN_modulation[-1].bias, 0) - nn.init.constant_(self.final_layer.linear.weight, 0) - nn.init.constant_(self.final_layer.linear.bias, 0) - - def _prepare_mask(self, attention_mask: Optional[torch.Tensor], dtype: torch.dtype): - if attention_mask is not None: - assert attention_mask.ndim == 4 - attention_mask = attention_mask.to(dtype) - inverted_mask = 1.0 - attention_mask - return inverted_mask.masked_fill( - inverted_mask.to(torch.bool), torch.finfo(dtype).min - ) - return attention_mask - - def enable_gradient_checkpointing(self): - self.grad_checkpointing = True - - def disable_gradient_checkpointing(self): - self.grad_checkpointing = False - - def forward( - self, - video_latent_states, - t, - text_latent_states=None, - attention_mask=None, - ): - """ - video_latent_states: [B, S, C, P, P] - t: [B] - text_latent_states: [B, H] if model_arch == "adaln" else [B, S, H] - attention_mask: [B, 1, Sv, St] or [B, 1, Sv, Sv] - """ - video_latent_states = self.video_embedder(video_latent_states) - text_len = text_latent_states.shape[1] - text_latent_states = self.text_embedder(text_latent_states) - if self.model_arch == "token-concat": - video_latent_states = torch.cat( - [text_latent_states, video_latent_states], dim=1 - ) - text_latent_states = None - pos_embed = self.pos_embed(video_latent_states) - video_latent_states = video_latent_states + pos_embed - t = self.t_embedder(t) # (N, D) - if self.model_arch == "adaln": - t = t + text_latent_states - text_latent_states = None - attention_mask = self._prepare_mask(attention_mask, video_latent_states.dtype) - - if self.seq_parallel_group is not None and self.seq_parallel_size > 1: - assert video_latent_states.shape[1] % self.seq_parallel_size == 0 - video_latent_states = split_seq( - video_latent_states, self.seq_parallel_size, self.seq_parallel_rank - ) - if text_latent_states is not None: - assert text_latent_states.shape[1] % self.seq_parallel_size == 0 - text_latent_states = split_seq( - text_latent_states, self.seq_parallel_size, self.seq_parallel_rank - ) - - for block in self.blocks: - if self.grad_checkpointing and self.training: - video_latent_states = torch.utils.checkpoint.checkpoint( - block, - video_latent_states, - attention_mask, - t, - text_latent_states, - ) - else: - video_latent_states = block( - video_latent_states, attention_mask, t, text_latent_states - ) - - if self.seq_parallel_group is not None and self.seq_parallel_size > 1: - video_latent_states = gather_seq( - video_latent_states, - self.seq_parallel_size, - self.seq_parallel_rank, - self.seq_parallel_group, - ) - - if self.model_arch == "token-concat": - video_latent_states = video_latent_states[:, text_len:] - video_latent_states = self.final_layer(video_latent_states, t) - return video_latent_states - - def forward_with_cfg( - self, x, t, text_latent_states, cfg_scale, attention_mask=None - ): - """ - Forward pass of DiT, but also batches the unconditional forward pass for classifier-free guidance. - """ - # https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb - half = x[: len(x) // 2] - combined = torch.cat([half, half], dim=0) - model_out = self.forward( - combined, t, text_latent_states, attention_mask=attention_mask - ) - # For exact reproducibility reasons, we apply classifier-free guidance on only - # three channels by default. The standard approach to cfg applies it to all channels. - # This can be done by uncommenting the following line and commenting-out the line following that. - # eps, rest = model_out[:, :self.in_channels], model_out[:, self.in_channels:] - if self.learn_sigma: - c = model_out.shape[2] - assert c == 2 * self.in_channels - eps, rest = model_out.chunk(2, dim=2) - else: - eps = model_out - cond_eps, uncond_eps = torch.split(eps, len(eps) // 2, dim=0) - half_eps = uncond_eps + cfg_scale * (cond_eps - uncond_eps) - eps = torch.cat([half_eps, half_eps], dim=0) - if self.learn_sigma: - return torch.cat([eps, rest], dim=2) - return eps - - -################################################################################# -# Sine/Cosine Positional Embedding Functions # -################################################################################# -# https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py - - -def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=0): - """ - grid_size: int of the grid height and width - return: - pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token) - """ - grid_h = np.arange(grid_size, dtype=np.float32) - grid_w = np.arange(grid_size, dtype=np.float32) - grid = np.meshgrid(grid_w, grid_h) # here w goes first - grid = np.stack(grid, axis=0) - - grid = grid.reshape([2, 1, grid_size, grid_size]) - pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid) - if cls_token and extra_tokens > 0: - pos_embed = np.concatenate( - [np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0 - ) - return pos_embed - - -def get_2d_sincos_pos_embed_from_grid(embed_dim, grid): - assert embed_dim % 2 == 0 - - # use half of dimensions to encode grid_h - emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2) - emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2) - - emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D) - return emb - - -def get_1d_sincos_pos_embed_from_grid(embed_dim, pos): - """ - embed_dim: output dimension for each position - pos: a list of positions to be encoded: size (M,) - out: (M, D) - """ - assert embed_dim % 2 == 0 - omega = np.arange(embed_dim // 2, dtype=np.float64) - omega /= embed_dim / 2.0 - omega = 1.0 / 10000**omega # (D/2,) - - pos = pos.reshape(-1) # (M,) - out = np.einsum("m,d->md", pos, omega) # (M, D/2), outer product - - emb_sin = np.sin(out) # (M, D/2) - emb_cos = np.cos(out) # (M, D/2) - - emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D) - return emb - - -################################################################################# -# DiT Configs # -################################################################################# - - -def DiT_XL_2(**kwargs): - return DiT(depth=28, hidden_size=1152, patch_size=2, num_heads=16, **kwargs) - - -def DiT_XL_4(**kwargs): - return DiT(depth=28, hidden_size=1152, patch_size=4, num_heads=16, **kwargs) - - -def DiT_XL_8(**kwargs): - return DiT(depth=28, hidden_size=1152, patch_size=8, num_heads=16, **kwargs) - - -def DiT_L_2(**kwargs): - return DiT(depth=24, hidden_size=1024, patch_size=2, num_heads=16, **kwargs) - - -def DiT_L_4(**kwargs): - return DiT(depth=24, hidden_size=1024, patch_size=4, num_heads=16, **kwargs) - - -def DiT_L_8(**kwargs): - return DiT(depth=24, hidden_size=1024, patch_size=8, num_heads=16, **kwargs) - - -def DiT_B_2(**kwargs): - return DiT(depth=12, hidden_size=768, patch_size=2, num_heads=12, **kwargs) - - -def DiT_B_4(**kwargs): - return DiT(depth=12, hidden_size=768, patch_size=4, num_heads=12, **kwargs) - - -def DiT_B_8(**kwargs): - return DiT(depth=12, hidden_size=768, patch_size=8, num_heads=12, **kwargs) - - -def DiT_S_2(**kwargs): - return DiT(depth=12, hidden_size=384, patch_size=2, num_heads=6, **kwargs) - - -def DiT_S_4(**kwargs): - return DiT(depth=12, hidden_size=384, patch_size=4, num_heads=6, **kwargs) - - -def DiT_S_8(**kwargs): - return DiT(depth=12, hidden_size=384, patch_size=8, num_heads=6, **kwargs) - - -DiT_models = { - "DiT-XL/2": DiT_XL_2, - "DiT-XL/4": DiT_XL_4, - "DiT-XL/8": DiT_XL_8, - "DiT-L/2": DiT_L_2, - "DiT-L/4": DiT_L_4, - "DiT-L/8": DiT_L_8, - "DiT-B/2": DiT_B_2, - "DiT-B/4": DiT_B_4, - "DiT-B/8": DiT_B_8, - "DiT-S/2": DiT_S_2, - "DiT-S/4": DiT_S_4, - "DiT-S/8": DiT_S_8, -} diff --git a/open_sora/modeling/latte/__init__.py b/open_sora/modeling/latte/__init__.py deleted file mode 100644 index 159db0f3..00000000 --- a/open_sora/modeling/latte/__init__.py +++ /dev/null @@ -1,3 +0,0 @@ -from .latte_t2v import LatteT2V - -__all__ = ["LatteT2V"] diff --git a/open_sora/modeling/latte/latte_t2v.py b/open_sora/modeling/latte/latte_t2v.py deleted file mode 100644 index 7c4de50c..00000000 --- a/open_sora/modeling/latte/latte_t2v.py +++ /dev/null @@ -1,1003 +0,0 @@ -import json -import os -from dataclasses import dataclass -from typing import Any, Dict, Optional, Tuple - -import torch -import torch.nn.functional as F -from diffusers.configuration_utils import ConfigMixin, register_to_config -from diffusers.models.activations import GEGLU, GELU, ApproximateGELU -from diffusers.models.attention import BasicTransformerBlock -from diffusers.models.attention_processor import Attention -from diffusers.models.embeddings import ( - CaptionProjection, - CombinedTimestepSizeEmbeddings, - ImagePositionalEmbeddings, - PatchEmbed, - SinusoidalPositionalEmbedding, - get_1d_sincos_pos_embed_from_grid, -) -from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear -from diffusers.models.modeling_utils import ModelMixin -from diffusers.models.normalization import AdaLayerNorm, AdaLayerNormZero -from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate -from diffusers.utils.torch_utils import maybe_allow_in_graph -from einops import rearrange, repeat -from torch import nn - - -@maybe_allow_in_graph -class GatedSelfAttentionDense(nn.Module): - r""" - A gated self-attention dense layer that combines visual features and object features. - - Parameters: - query_dim (`int`): The number of channels in the query. - context_dim (`int`): The number of channels in the context. - n_heads (`int`): The number of heads to use for attention. - d_head (`int`): The number of channels in each head. - """ - - def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int): - super().__init__() - - # we need a linear projection since we need cat visual feature and obj feature - self.linear = nn.Linear(context_dim, query_dim) - - self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head) - self.ff = FeedForward(query_dim, activation_fn="geglu") - - self.norm1 = nn.LayerNorm(query_dim) - self.norm2 = nn.LayerNorm(query_dim) - - self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0))) - self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0))) - - self.enabled = True - - def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor: - if not self.enabled: - return x - - n_visual = x.shape[1] - objs = self.linear(objs) - - x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :] - x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x)) - - return x - - -class FeedForward(nn.Module): - r""" - A feed-forward layer. - - Parameters: - dim (`int`): The number of channels in the input. - dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`. - mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension. - dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. - activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. - final_dropout (`bool` *optional*, defaults to False): Apply a final dropout. - """ - - def __init__( - self, - dim: int, - dim_out: Optional[int] = None, - mult: int = 4, - dropout: float = 0.0, - activation_fn: str = "geglu", - final_dropout: bool = False, - ): - super().__init__() - inner_dim = int(dim * mult) - dim_out = dim_out if dim_out is not None else dim - linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear - - if activation_fn == "gelu": - act_fn = GELU(dim, inner_dim) - if activation_fn == "gelu-approximate": - act_fn = GELU(dim, inner_dim, approximate="tanh") - elif activation_fn == "geglu": - act_fn = GEGLU(dim, inner_dim) - elif activation_fn == "geglu-approximate": - act_fn = ApproximateGELU(dim, inner_dim) - - self.net = nn.ModuleList([]) - # project in - self.net.append(act_fn) - # project dropout - self.net.append(nn.Dropout(dropout)) - # project out - self.net.append(linear_cls(inner_dim, dim_out)) - # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout - if final_dropout: - self.net.append(nn.Dropout(dropout)) - - def forward(self, hidden_states: torch.Tensor, scale: float = 1.0) -> torch.Tensor: - compatible_cls = (GEGLU,) if USE_PEFT_BACKEND else (GEGLU, LoRACompatibleLinear) - for module in self.net: - if isinstance(module, compatible_cls): - hidden_states = module(hidden_states, scale) - else: - hidden_states = module(hidden_states) - return hidden_states - - -@maybe_allow_in_graph -class BasicTransformerBlock_(nn.Module): - r""" - A basic Transformer block. - - Parameters: - dim (`int`): The number of channels in the input and output. - num_attention_heads (`int`): The number of heads to use for multi-head attention. - attention_head_dim (`int`): The number of channels in each head. - dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. - cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. - activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. - num_embeds_ada_norm (: - obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`. - attention_bias (: - obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter. - only_cross_attention (`bool`, *optional*): - Whether to use only cross-attention layers. In this case two cross attention layers are used. - double_self_attention (`bool`, *optional*): - Whether to use two self-attention layers. In this case no cross attention layers are used. - upcast_attention (`bool`, *optional*): - Whether to upcast the attention computation to float32. This is useful for mixed precision training. - norm_elementwise_affine (`bool`, *optional*, defaults to `True`): - Whether to use learnable elementwise affine parameters for normalization. - norm_type (`str`, *optional*, defaults to `"layer_norm"`): - The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`. - final_dropout (`bool` *optional*, defaults to False): - Whether to apply a final dropout after the last feed-forward layer. - attention_type (`str`, *optional*, defaults to `"default"`): - The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`. - positional_embeddings (`str`, *optional*, defaults to `None`): - The type of positional embeddings to apply to. - num_positional_embeddings (`int`, *optional*, defaults to `None`): - The maximum number of positional embeddings to apply. - """ - - def __init__( - self, - dim: int, - num_attention_heads: int, - attention_head_dim: int, - dropout=0.0, - cross_attention_dim: Optional[int] = None, - activation_fn: str = "geglu", - num_embeds_ada_norm: Optional[int] = None, - attention_bias: bool = False, - only_cross_attention: bool = False, - double_self_attention: bool = False, - upcast_attention: bool = False, - norm_elementwise_affine: bool = True, - norm_type: str = "layer_norm", # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single' - norm_eps: float = 1e-5, - final_dropout: bool = False, - attention_type: str = "default", - positional_embeddings: Optional[str] = None, - num_positional_embeddings: Optional[int] = None, - ): - super().__init__() - self.only_cross_attention = only_cross_attention - - self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero" - self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm" - self.use_ada_layer_norm_single = norm_type == "ada_norm_single" - self.use_layer_norm = norm_type == "layer_norm" - - if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None: - raise ValueError( - f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to" - f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}." - ) - - if positional_embeddings and (num_positional_embeddings is None): - raise ValueError( - "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined." - ) - - if positional_embeddings == "sinusoidal": - self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings) - else: - self.pos_embed = None - - # Define 3 blocks. Each block has its own normalization layer. - # 1. Self-Attn - if self.use_ada_layer_norm: - self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) - elif self.use_ada_layer_norm_zero: - self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm) - else: - self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps) - - self.attn1 = Attention( - query_dim=dim, - heads=num_attention_heads, - dim_head=attention_head_dim, - dropout=dropout, - bias=attention_bias, - cross_attention_dim=cross_attention_dim if only_cross_attention else None, - upcast_attention=upcast_attention, - ) - - # # 2. Cross-Attn - # if cross_attention_dim is not None or double_self_attention: - # # We currently only use AdaLayerNormZero for self attention where there will only be one attention block. - # # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during - # # the second cross attention block. - # self.norm2 = ( - # AdaLayerNorm(dim, num_embeds_ada_norm) - # if self.use_ada_layer_norm - # else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps) - # ) - # self.attn2 = Attention( - # query_dim=dim, - # cross_attention_dim=cross_attention_dim if not double_self_attention else None, - # heads=num_attention_heads, - # dim_head=attention_head_dim, - # dropout=dropout, - # bias=attention_bias, - # upcast_attention=upcast_attention, - # ) # is self-attn if encoder_hidden_states is none - # else: - # self.norm2 = None - # self.attn2 = None - - # 3. Feed-forward - # if not self.use_ada_layer_norm_single: - # self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps) - self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps) - - self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout) - - # 4. Fuser - if attention_type == "gated" or attention_type == "gated-text-image": - self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim) - - # 5. Scale-shift for PixArt-Alpha. - if self.use_ada_layer_norm_single: - self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5) - - # let chunk size default to None - self._chunk_size = None - self._chunk_dim = 0 - - def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int): - # Sets chunk feed-forward - self._chunk_size = chunk_size - self._chunk_dim = dim - - def forward( - self, - hidden_states: torch.FloatTensor, - attention_mask: Optional[torch.FloatTensor] = None, - encoder_hidden_states: Optional[torch.FloatTensor] = None, - encoder_attention_mask: Optional[torch.FloatTensor] = None, - timestep: Optional[torch.LongTensor] = None, - cross_attention_kwargs: Dict[str, Any] = None, - class_labels: Optional[torch.LongTensor] = None, - ) -> torch.FloatTensor: - # Notice that normalization is always applied before the real computation in the following blocks. - # 0. Self-Attention - batch_size = hidden_states.shape[0] - - if self.use_ada_layer_norm: - norm_hidden_states = self.norm1(hidden_states, timestep) - elif self.use_ada_layer_norm_zero: - norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1( - hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype - ) - elif self.use_layer_norm: - norm_hidden_states = self.norm1(hidden_states) - elif self.use_ada_layer_norm_single: - shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ( - self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1) - ).chunk(6, dim=1) - norm_hidden_states = self.norm1(hidden_states) - norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa - norm_hidden_states = norm_hidden_states.squeeze(1) - else: - raise ValueError("Incorrect norm used") - - if self.pos_embed is not None: - norm_hidden_states = self.pos_embed(norm_hidden_states) - - # 1. Retrieve lora scale. - lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 - - # 2. Prepare GLIGEN inputs - cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {} - gligen_kwargs = cross_attention_kwargs.pop("gligen", None) - - attn_output = self.attn1( - norm_hidden_states, - encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None, - attention_mask=attention_mask, - **cross_attention_kwargs, - ) - if self.use_ada_layer_norm_zero: - attn_output = gate_msa.unsqueeze(1) * attn_output - elif self.use_ada_layer_norm_single: - attn_output = gate_msa * attn_output - - hidden_states = attn_output + hidden_states - if hidden_states.ndim == 4: - hidden_states = hidden_states.squeeze(1) - - # 2.5 GLIGEN Control - if gligen_kwargs is not None: - hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"]) - - # # 3. Cross-Attention - # if self.attn2 is not None: - # if self.use_ada_layer_norm: - # norm_hidden_states = self.norm2(hidden_states, timestep) - # elif self.use_ada_layer_norm_zero or self.use_layer_norm: - # norm_hidden_states = self.norm2(hidden_states) - # elif self.use_ada_layer_norm_single: - # # For PixArt norm2 isn't applied here: - # # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103 - # norm_hidden_states = hidden_states - # else: - # raise ValueError("Incorrect norm") - - # if self.pos_embed is not None and self.use_ada_layer_norm_single is False: - # norm_hidden_states = self.pos_embed(norm_hidden_states) - - # attn_output = self.attn2( - # norm_hidden_states, - # encoder_hidden_states=encoder_hidden_states, - # attention_mask=encoder_attention_mask, - # **cross_attention_kwargs, - # ) - # hidden_states = attn_output + hidden_states - - # 4. Feed-forward - # if not self.use_ada_layer_norm_single: - # norm_hidden_states = self.norm3(hidden_states) - - if self.use_ada_layer_norm_zero: - norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None] - - if self.use_ada_layer_norm_single: - # norm_hidden_states = self.norm2(hidden_states) - norm_hidden_states = self.norm3(hidden_states) - norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp - - if self._chunk_size is not None: - # "feed_forward_chunk_size" can be used to save memory - if norm_hidden_states.shape[self._chunk_dim] % self._chunk_size != 0: - raise ValueError( - f"`hidden_states` dimension to be chunked: {norm_hidden_states.shape[self._chunk_dim]} has to be divisible by chunk size: {self._chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`." - ) - - num_chunks = norm_hidden_states.shape[self._chunk_dim] // self._chunk_size - ff_output = torch.cat( - [ - self.ff(hid_slice, scale=lora_scale) - for hid_slice in norm_hidden_states.chunk(num_chunks, dim=self._chunk_dim) - ], - dim=self._chunk_dim, - ) - else: - ff_output = self.ff(norm_hidden_states, scale=lora_scale) - - if self.use_ada_layer_norm_zero: - ff_output = gate_mlp.unsqueeze(1) * ff_output - elif self.use_ada_layer_norm_single: - ff_output = gate_mlp * ff_output - - hidden_states = ff_output + hidden_states - if hidden_states.ndim == 4: - hidden_states = hidden_states.squeeze(1) - - return hidden_states - - -class AdaLayerNormSingle(nn.Module): - r""" - Norm layer adaptive layer norm single (adaLN-single). - - As proposed in PixArt-Alpha (see: https://arxiv.org/abs/2310.00426; Section 2.3). - - Parameters: - embedding_dim (`int`): The size of each embedding vector. - use_additional_conditions (`bool`): To use additional conditions for normalization or not. - """ - - def __init__(self, embedding_dim: int, use_additional_conditions: bool = False): - super().__init__() - - self.emb = CombinedTimestepSizeEmbeddings( - embedding_dim, size_emb_dim=embedding_dim // 3, use_additional_conditions=use_additional_conditions - ) - - self.silu = nn.SiLU() - self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=True) - - def forward( - self, - timestep: torch.Tensor, - added_cond_kwargs: Dict[str, torch.Tensor] = None, - batch_size: int = None, - hidden_dtype: Optional[torch.dtype] = None, - ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: - # No modulation happening here. - embedded_timestep = self.emb( - timestep, batch_size=batch_size, hidden_dtype=hidden_dtype, resolution=None, aspect_ratio=None - ) - return self.linear(self.silu(embedded_timestep)), embedded_timestep - - -@dataclass -class Transformer3DModelOutput(BaseOutput): - """ - The output of [`Transformer2DModel`]. - - Args: - sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete): - The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability - distributions for the unnoised latent pixels. - """ - - sample: torch.FloatTensor - - -class LatteT2V(ModelMixin, ConfigMixin): - _supports_gradient_checkpointing = True - - """ - A 2D Transformer model for image-like data. - - Parameters: - num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention. - attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head. - in_channels (`int`, *optional*): - The number of channels in the input and output (specify if the input is **continuous**). - num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use. - dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. - cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use. - sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**). - This is fixed during training since it is used to learn a number of position embeddings. - num_vector_embeds (`int`, *optional*): - The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**). - Includes the class for the masked latent pixel. - activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward. - num_embeds_ada_norm ( `int`, *optional*): - The number of diffusion steps used during training. Pass if at least one of the norm_layers is - `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are - added to the hidden states. - - During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`. - attention_bias (`bool`, *optional*): - Configure if the `TransformerBlocks` attention should contain a bias parameter. - """ - - @register_to_config - def __init__( - self, - num_attention_heads: int = 16, - attention_head_dim: int = 88, - in_channels: Optional[int] = None, - out_channels: Optional[int] = None, - num_layers: int = 1, - dropout: float = 0.0, - norm_num_groups: int = 32, - cross_attention_dim: Optional[int] = None, - attention_bias: bool = False, - sample_size: Optional[int] = None, - num_vector_embeds: Optional[int] = None, - patch_size: Optional[int] = None, - activation_fn: str = "geglu", - num_embeds_ada_norm: Optional[int] = None, - use_linear_projection: bool = False, - only_cross_attention: bool = False, - double_self_attention: bool = False, - upcast_attention: bool = False, - norm_type: str = "layer_norm", - norm_elementwise_affine: bool = True, - norm_eps: float = 1e-5, - attention_type: str = "default", - caption_channels: int = None, - video_length: int = 16, - ): - super().__init__() - self.use_linear_projection = use_linear_projection - self.num_attention_heads = num_attention_heads - self.attention_head_dim = attention_head_dim - inner_dim = num_attention_heads * attention_head_dim - self.video_length = video_length - - conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv - linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear - - # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)` - # Define whether input is continuous or discrete depending on configuration - self.is_input_continuous = (in_channels is not None) and (patch_size is None) - self.is_input_vectorized = num_vector_embeds is not None - self.is_input_patches = in_channels is not None and patch_size is not None - - if norm_type == "layer_norm" and num_embeds_ada_norm is not None: - deprecation_message = ( - f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or" - " incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config." - " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect" - " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it" - " would be very nice if you could open a Pull request for the `transformer/config.json` file" - ) - deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False) - norm_type = "ada_norm" - - if self.is_input_continuous and self.is_input_vectorized: - raise ValueError( - f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make" - " sure that either `in_channels` or `num_vector_embeds` is None." - ) - elif self.is_input_vectorized and self.is_input_patches: - raise ValueError( - f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make" - " sure that either `num_vector_embeds` or `num_patches` is None." - ) - elif not self.is_input_continuous and not self.is_input_vectorized and not self.is_input_patches: - raise ValueError( - f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:" - f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None." - ) - - # 2. Define input layers - if self.is_input_continuous: - self.in_channels = in_channels - - self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True) - if use_linear_projection: - self.proj_in = linear_cls(in_channels, inner_dim) - else: - self.proj_in = conv_cls(in_channels, inner_dim, kernel_size=1, stride=1, padding=0) - elif self.is_input_vectorized: - assert sample_size is not None, "Transformer2DModel over discrete input must provide sample_size" - assert num_vector_embeds is not None, "Transformer2DModel over discrete input must provide num_embed" - - self.height = sample_size - self.width = sample_size - self.num_vector_embeds = num_vector_embeds - self.num_latent_pixels = self.height * self.width - - self.latent_image_embedding = ImagePositionalEmbeddings( - num_embed=num_vector_embeds, embed_dim=inner_dim, height=self.height, width=self.width - ) - elif self.is_input_patches: - assert sample_size is not None, "Transformer2DModel over patched input must provide sample_size" - - self.height = sample_size - self.width = sample_size - - self.patch_size = patch_size - interpolation_scale = self.config.sample_size // 64 # => 64 (= 512 pixart) has interpolation scale 1 - interpolation_scale = max(interpolation_scale, 1) - self.pos_embed = PatchEmbed( - height=sample_size, - width=sample_size, - patch_size=patch_size, - in_channels=in_channels, - embed_dim=inner_dim, - interpolation_scale=interpolation_scale, - ) - - # 3. Define transformers blocks - self.transformer_blocks = nn.ModuleList( - [ - BasicTransformerBlock( - inner_dim, - num_attention_heads, - attention_head_dim, - dropout=dropout, - cross_attention_dim=cross_attention_dim, - activation_fn=activation_fn, - num_embeds_ada_norm=num_embeds_ada_norm, - attention_bias=attention_bias, - only_cross_attention=only_cross_attention, - double_self_attention=double_self_attention, - upcast_attention=upcast_attention, - norm_type=norm_type, - norm_elementwise_affine=norm_elementwise_affine, - norm_eps=norm_eps, - attention_type=attention_type, - ) - for d in range(num_layers) - ] - ) - - # Define temporal transformers blocks - self.temporal_transformer_blocks = nn.ModuleList( - [ - BasicTransformerBlock_( # one attention - inner_dim, - num_attention_heads, # num_attention_heads - attention_head_dim, # attention_head_dim 72 - dropout=dropout, - cross_attention_dim=None, - activation_fn=activation_fn, - num_embeds_ada_norm=num_embeds_ada_norm, - attention_bias=attention_bias, - only_cross_attention=only_cross_attention, - double_self_attention=False, - upcast_attention=upcast_attention, - norm_type=norm_type, - norm_elementwise_affine=norm_elementwise_affine, - norm_eps=norm_eps, - attention_type=attention_type, - ) - for d in range(num_layers) - ] - ) - - # 4. Define output layers - self.out_channels = in_channels if out_channels is None else out_channels - if self.is_input_continuous: - # TODO: should use out_channels for continuous projections - if use_linear_projection: - self.proj_out = linear_cls(inner_dim, in_channels) - else: - self.proj_out = conv_cls(inner_dim, in_channels, kernel_size=1, stride=1, padding=0) - elif self.is_input_vectorized: - self.norm_out = nn.LayerNorm(inner_dim) - self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1) - elif self.is_input_patches and norm_type != "ada_norm_single": - self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6) - self.proj_out_1 = nn.Linear(inner_dim, 2 * inner_dim) - self.proj_out_2 = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels) - elif self.is_input_patches and norm_type == "ada_norm_single": - self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6) - self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5) - self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels) - - # 5. PixArt-Alpha blocks. - self.adaln_single = None - self.use_additional_conditions = False - if norm_type == "ada_norm_single": - self.use_additional_conditions = self.config.sample_size == 128 # False, 128 -> 1024 - # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use - # additional conditions until we find better name - self.adaln_single = AdaLayerNormSingle(inner_dim, use_additional_conditions=self.use_additional_conditions) - - self.caption_projection = None - if caption_channels is not None: - self.caption_projection = CaptionProjection(in_features=caption_channels, hidden_size=inner_dim) - - self.gradient_checkpointing = False - - # define temporal positional embedding - temp_pos_embed = self.get_1d_sincos_temp_embed(inner_dim, video_length) # 1152 hidden size - self.register_buffer("temp_pos_embed", torch.from_numpy(temp_pos_embed).float().unsqueeze(0), persistent=False) - - def _set_gradient_checkpointing(self, module, value=False): - self.gradient_checkpointing = value - - def forward( - self, - hidden_states: torch.Tensor, - timestep: Optional[torch.LongTensor] = None, - encoder_hidden_states: Optional[torch.Tensor] = None, - added_cond_kwargs: Dict[str, torch.Tensor] = None, - class_labels: Optional[torch.LongTensor] = None, - cross_attention_kwargs: Dict[str, Any] = None, - attention_mask: Optional[torch.Tensor] = None, - encoder_attention_mask: Optional[torch.Tensor] = None, - use_image_num: int = 0, - enable_temporal_attentions: bool = True, - return_dict: bool = True, - ): - """ - The [`Transformer2DModel`] forward method. - - Args: - hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, frame, channel, height, width)` if continuous): - Input `hidden_states`. - encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*): - Conditional embeddings for cross attention layer. If not given, cross-attention defaults to - self-attention. - timestep ( `torch.LongTensor`, *optional*): - Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`. - class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*): - Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in - `AdaLayerZeroNorm`. - cross_attention_kwargs ( `Dict[str, Any]`, *optional*): - A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under - `self.processor` in - [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). - attention_mask ( `torch.Tensor`, *optional*): - An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask - is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large - negative values to the attention scores corresponding to "discard" tokens. - encoder_attention_mask ( `torch.Tensor`, *optional*): - Cross-attention mask applied to `encoder_hidden_states`. Two formats supported: - - * Mask `(batch, sequence_length)` True = keep, False = discard. - * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard. - - If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format - above. This bias will be added to the cross-attention scores. - return_dict (`bool`, *optional*, defaults to `True`): - Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain - tuple. - - Returns: - If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a - `tuple` where the first element is the sample tensor. - """ - input_batch_size, c, frame, h, w = hidden_states.shape - frame = frame - use_image_num - hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w").contiguous() - - # ensure attention_mask is a bias, and give it a singleton query_tokens dimension. - # we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward. - # we can tell by counting dims; if ndim == 2: it's a mask rather than a bias. - # expects mask of shape: - # [batch, key_tokens] - # adds singleton query_tokens dimension: - # [batch, 1, key_tokens] - # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes: - # [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn) - # [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn) - if attention_mask is not None and attention_mask.ndim == 2: - # assume that mask is expressed as: - # (1 = keep, 0 = discard) - # convert mask into a bias that can be added to attention scores: - # (keep = +0, discard = -10000.0) - attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0 - attention_mask = attention_mask.unsqueeze(1) - - # convert encoder_attention_mask to a bias the same way we do for attention_mask - if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2: # ndim == 2 means no image joint - encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0 - encoder_attention_mask = encoder_attention_mask.unsqueeze(1) - encoder_attention_mask = repeat(encoder_attention_mask, "b 1 l -> (b f) 1 l", f=frame).contiguous() - elif encoder_attention_mask is not None and encoder_attention_mask.ndim == 3: # ndim == 3 means image joint - encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0 - encoder_attention_mask_video = encoder_attention_mask[:, :1, ...] - encoder_attention_mask_video = repeat( - encoder_attention_mask_video, "b 1 l -> b (1 f) l", f=frame - ).contiguous() - encoder_attention_mask_image = encoder_attention_mask[:, 1:, ...] - encoder_attention_mask = torch.cat([encoder_attention_mask_video, encoder_attention_mask_image], dim=1) - encoder_attention_mask = rearrange(encoder_attention_mask, "b n l -> (b n) l").contiguous().unsqueeze(1) - - # Retrieve lora scale. - cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 - - # 1. Input - if self.is_input_patches: # here - height, width = hidden_states.shape[-2] // self.patch_size, hidden_states.shape[-1] // self.patch_size - num_patches = height * width - - hidden_states = self.pos_embed(hidden_states) # already add positional embeddings - - if self.adaln_single is not None: - if self.use_additional_conditions and added_cond_kwargs is None: - raise ValueError( - "`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`." - ) - # batch_size = hidden_states.shape[0] - batch_size = input_batch_size - timestep, embedded_timestep = self.adaln_single( - timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype - ) - - # 2. Blocks - if self.caption_projection is not None: - batch_size = hidden_states.shape[0] - encoder_hidden_states = self.caption_projection(encoder_hidden_states) # 3 120 1152 - - if use_image_num != 0 and self.training: - encoder_hidden_states_video = encoder_hidden_states[:, :1, ...] - encoder_hidden_states_video = repeat( - encoder_hidden_states_video, "b 1 t d -> b (1 f) t d", f=frame - ).contiguous() - encoder_hidden_states_image = encoder_hidden_states[:, 1:, ...] - encoder_hidden_states = torch.cat([encoder_hidden_states_video, encoder_hidden_states_image], dim=1) - encoder_hidden_states_spatial = rearrange(encoder_hidden_states, "b f t d -> (b f) t d").contiguous() - else: - encoder_hidden_states_spatial = repeat( - encoder_hidden_states, "b t d -> (b f) t d", f=frame - ).contiguous() - - # prepare timesteps for spatial and temporal block - timestep_spatial = repeat(timestep, "b d -> (b f) d", f=frame + use_image_num).contiguous() - timestep_temp = repeat(timestep, "b d -> (b p) d", p=num_patches).contiguous() - - for i, (spatial_block, temp_block) in enumerate(zip(self.transformer_blocks, self.temporal_transformer_blocks)): - if self.training and self.gradient_checkpointing: - hidden_states = torch.utils.checkpoint.checkpoint( - spatial_block, - hidden_states, - attention_mask, - encoder_hidden_states_spatial, - encoder_attention_mask, - timestep_spatial, - cross_attention_kwargs, - class_labels, - use_reentrant=False, - ) - - if enable_temporal_attentions: - hidden_states = rearrange(hidden_states, "(b f) t d -> (b t) f d", b=input_batch_size).contiguous() - - if use_image_num != 0: # image-video joitn training - hidden_states_video = hidden_states[:, :frame, ...] - hidden_states_image = hidden_states[:, frame:, ...] - - if i == 0: - hidden_states_video = hidden_states_video + self.temp_pos_embed - - hidden_states_video = torch.utils.checkpoint.checkpoint( - temp_block, - hidden_states_video, - None, # attention_mask - None, # encoder_hidden_states - None, # encoder_attention_mask - timestep_temp, - cross_attention_kwargs, - class_labels, - use_reentrant=False, - ) - - hidden_states = torch.cat([hidden_states_video, hidden_states_image], dim=1) - hidden_states = rearrange( - hidden_states, "(b t) f d -> (b f) t d", b=input_batch_size - ).contiguous() - - else: - if i == 0: - hidden_states = hidden_states + self.temp_pos_embed - - hidden_states = torch.utils.checkpoint.checkpoint( - temp_block, - hidden_states, - None, # attention_mask - None, # encoder_hidden_states - None, # encoder_attention_mask - timestep_temp, - cross_attention_kwargs, - class_labels, - use_reentrant=False, - ) - - hidden_states = rearrange( - hidden_states, "(b t) f d -> (b f) t d", b=input_batch_size - ).contiguous() - else: - hidden_states = spatial_block( - hidden_states, - attention_mask, - encoder_hidden_states_spatial, - encoder_attention_mask, - timestep_spatial, - cross_attention_kwargs, - class_labels, - ) - - if enable_temporal_attentions: - hidden_states = rearrange(hidden_states, "(b f) t d -> (b t) f d", b=input_batch_size).contiguous() - - if use_image_num != 0 and self.training: - hidden_states_video = hidden_states[:, :frame, ...] - hidden_states_image = hidden_states[:, frame:, ...] - - hidden_states_video = temp_block( - hidden_states_video, - None, # attention_mask - None, # encoder_hidden_states - None, # encoder_attention_mask - timestep_temp, - cross_attention_kwargs, - class_labels, - ) - - hidden_states = torch.cat([hidden_states_video, hidden_states_image], dim=1) - hidden_states = rearrange( - hidden_states, "(b t) f d -> (b f) t d", b=input_batch_size - ).contiguous() - - else: - if i == 0: - hidden_states = hidden_states + self.temp_pos_embed - - hidden_states = temp_block( - hidden_states, - None, # attention_mask - None, # encoder_hidden_states - None, # encoder_attention_mask - timestep_temp, - cross_attention_kwargs, - class_labels, - ) - - hidden_states = rearrange( - hidden_states, "(b t) f d -> (b f) t d", b=input_batch_size - ).contiguous() - - if self.is_input_patches: - if self.config.norm_type != "ada_norm_single": - conditioning = self.transformer_blocks[0].norm1.emb( - timestep, class_labels, hidden_dtype=hidden_states.dtype - ) - shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1) - hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None] - hidden_states = self.proj_out_2(hidden_states) - elif self.config.norm_type == "ada_norm_single": - embedded_timestep = repeat(embedded_timestep, "b d -> (b f) d", f=frame + use_image_num).contiguous() - shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1) - hidden_states = self.norm_out(hidden_states) - # Modulation - hidden_states = hidden_states * (1 + scale) + shift - hidden_states = self.proj_out(hidden_states) - - # unpatchify - if self.adaln_single is None: - height = width = int(hidden_states.shape[1] ** 0.5) - hidden_states = hidden_states.reshape( - shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels) - ) - hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states) - output = hidden_states.reshape( - shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size) - ) - output = rearrange(output, "(b f) c h w -> b c f h w", b=input_batch_size).contiguous() - - if not return_dict: - return (output,) - - return Transformer3DModelOutput(sample=output) - - def get_1d_sincos_temp_embed(self, embed_dim, length): - pos = torch.arange(0, length).unsqueeze(1) - return get_1d_sincos_pos_embed_from_grid(embed_dim, pos) - - @classmethod - def from_pretrained_2d(cls, pretrained_model_path, subfolder=None, **kwargs): - if subfolder is not None: - pretrained_model_path = os.path.join(pretrained_model_path, subfolder) - - config_file = os.path.join(pretrained_model_path, "config.json") - if not os.path.isfile(config_file): - raise RuntimeError(f"{config_file} does not exist") - with open(config_file, "r") as f: - config = json.load(f) - - model = cls.from_config(config, **kwargs) - - # model_files = [ - # os.path.join(pretrained_model_path, 'diffusion_pytorch_model.bin'), - # os.path.join(pretrained_model_path, 'diffusion_pytorch_model.safetensors') - # ] - - # model_file = None - - # for fp in model_files: - # if os.path.exists(fp): - # model_file = fp - - # if not model_file: - # raise RuntimeError(f"{model_file} does not exist") - - # if model_file.split(".")[-1] == "safetensors": - # from safetensors import safe_open - # state_dict = {} - # with safe_open(model_file, framework="pt", device="cpu") as f: - # for key in f.keys(): - # state_dict[key] = f.get_tensor(key) - # else: - # state_dict = torch.load(model_file, map_location="cpu") - - # for k, v in model.state_dict().items(): - # if 'temporal_transformer_blocks' in k: - # state_dict.update({k: v}) - - # model.load_state_dict(state_dict) - - return model diff --git a/open_sora/utils/comm.py b/open_sora/utils/comm.py deleted file mode 100644 index 9b947ff5..00000000 --- a/open_sora/utils/comm.py +++ /dev/null @@ -1,292 +0,0 @@ -from typing import Optional, Tuple - -import torch -import torch.distributed as dist -import torch.nn.functional as F -from colossalai.moe._operation import MoeInGradScaler, MoeOutGradScaler -from colossalai.shardformer.layer._operation import gather_forward_split_backward -from torch.distributed.distributed_c10d import get_global_rank - - -def _all_to_all( - input_: torch.Tensor, - world_size: int, - group: dist.ProcessGroup, - scatter_dim: int, - gather_dim: int, -): - input_list = [ - t.contiguous() for t in torch.tensor_split(input_, world_size, scatter_dim) - ] - output_list = [torch.empty_like(input_list[0]) for _ in range(world_size)] - dist.all_to_all(output_list, input_list, group=group) - return torch.cat(output_list, dim=gather_dim).contiguous() - - -class _AllToAll(torch.autograd.Function): - """All-to-all communication. - - Args: - input_: input matrix - process_group: communication group - scatter_dim: scatter dimension - gather_dim: gather dimension - """ - - @staticmethod - def forward(ctx, input_, process_group, scatter_dim, gather_dim): - ctx.process_group = process_group - ctx.scatter_dim = scatter_dim - ctx.gather_dim = gather_dim - ctx.world_size = dist.get_world_size(process_group) - return _all_to_all( - input_, ctx.world_size, process_group, scatter_dim, gather_dim - ) - - @staticmethod - def backward(ctx, grad_output): - return ( - _all_to_all( - grad_output, - ctx.world_size, - ctx.process_group, - ctx.gather_dim, - ctx.scatter_dim, - ), - None, - None, - None, - ) - - -def all_to_all( - input_: torch.Tensor, - process_group: dist.ProcessGroup, - scatter_dim: int = 2, - gather_dim: int = 1, -): - return _AllToAll.apply(input_, process_group, scatter_dim, gather_dim) - - -def split_seq(input_: torch.Tensor, sp_size: int, sp_rank: int, dim: int = 1): - """Split a tensor along sequence dimension. It will split input and divide grad by sp_size. - - Args: - input_ (torch.Tensor): The common shape is (bs, seq, *). - sp_size (int): Sequence parallel size. - sp_rank (int): Sequence parallel rank. - dim (int, optional): Sequence dimension. Defaults to 1. - """ - input_ = input_.chunk(sp_size, dim=dim)[sp_rank].clone() - return MoeOutGradScaler.apply(input_, sp_size) - - -def gather_seq( - input_: torch.Tensor, - sp_size: int, - sp_rank: int, - sp_group: dist.ProcessGroup, - dim: int = 1, -): - """Gather a tensor along sequence dimension. It will gather input and multiply grad by sp_size. - - Args: - input_ (torch.Tensor): The common shape is (bs, seq, *). - sp_size (int): Sequence parallel size. - sp_rank (int): Sequence parallel rank. - dim (int, optional): Sequence dimension. Defaults to 1. - """ - input_ = gather_forward_split_backward(input_, dim, sp_group) - return MoeInGradScaler.apply(input_, sp_size) - - -class AsyncAllGatherProjForTwo(torch.autograd.Function): - @staticmethod - def forward( - ctx, - hidden_states: torch.Tensor, - context: torch.Tensor, - q_proj_weight: torch.Tensor, - q_proj_bias: Optional[torch.Tensor], - k_proj_weight: torch.Tensor, - k_proj_bias: Optional[torch.Tensor], - v_proj_weight: torch.Tensor, - v_proj_bias: Optional[torch.Tensor], - dim: int, - process_group: dist.ProcessGroup, - sp_size: int, - sp_rank: int, - ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: - assert sp_size == 2 - ctx.process_group = process_group - ctx.sp_size = sp_size - ctx.sp_rank = sp_rank - ctx.dim = dim - - is_cross_attn = not (hidden_states is context) - ctx.is_cross_attn = is_cross_attn - - recv_hidden_states = torch.empty_like(hidden_states) - if is_cross_attn: - recv_context = torch.empty_like(context) - else: - recv_context = recv_hidden_states - - peer_global_rank = get_global_rank(process_group, 1 - sp_rank) - ops = [ - dist.P2POp(dist.isend, hidden_states, peer_global_rank, process_group), - dist.P2POp(dist.irecv, recv_hidden_states, peer_global_rank, process_group), - ] - if is_cross_attn: - ops.extend( - [ - dist.P2POp(dist.isend, context, peer_global_rank, process_group), - dist.P2POp( - dist.irecv, recv_context, peer_global_rank, process_group - ), - ] - ) - - reqs = dist.batch_isend_irecv(ops) - - # [B, S/P, H] - q = F.linear(hidden_states, q_proj_weight, q_proj_bias) - k = F.linear(context, k_proj_weight, k_proj_bias) - v = F.linear(context, v_proj_weight, v_proj_bias) - - for req in reqs: - req.wait() - - q_other = F.linear(recv_hidden_states, q_proj_weight, q_proj_bias) - k_other = F.linear(recv_context, k_proj_weight, k_proj_bias) - v_other = F.linear(recv_context, v_proj_weight, v_proj_bias) - - if sp_rank == 0: - q = torch.cat([q, q_other], dim=dim) - k = torch.cat([k, k_other], dim=dim) - v = torch.cat([v, v_other], dim=dim) - else: - q = torch.cat([q_other, q], dim=dim) - k = torch.cat([k_other, k], dim=dim) - v = torch.cat([v_other, v], dim=dim) - - ctx.save_for_backward( - hidden_states, - context, - q_proj_weight, - q_proj_bias, - k_proj_weight, - k_proj_bias, - v_proj_weight, - v_proj_bias, - recv_hidden_states, - recv_context, - ) - - return q, k, v - - @staticmethod - def backward( - ctx: torch.Any, q_grad: torch.Tensor, k_grad: torch.Tensor, v_grad: torch.Tensor - ) -> torch.Any: - ( - hidden_states, - context, - q_proj_weight, - q_proj_bias, - k_proj_weight, - k_proj_bias, - v_proj_weight, - v_proj_bias, - recv_hidden_states, - recv_context, - ) = ctx.saved_tensors - - # compute param grads - if ctx.sp_rank == 0: - hidden_states = torch.cat([hidden_states, recv_hidden_states], dim=ctx.dim) - if ctx.is_cross_attn: - context = torch.cat([context, recv_context], dim=ctx.dim) - else: - context = hidden_states - else: - hidden_states = torch.cat([recv_hidden_states, hidden_states], dim=ctx.dim) - if ctx.is_cross_attn: - context = torch.cat([recv_context, context], dim=ctx.dim) - else: - context = hidden_states - q_proj_weight_grad = q_grad.transpose(-1, -2).matmul(hidden_states).sum(dim=0) - q_proj_bias_grad = ( - q_grad.sum(dim=0).sum(dim=0) if q_proj_bias is not None else None - ) - k_proj_weight_grad = k_grad.transpose(-1, -2).matmul(context).sum(dim=0) - k_proj_bias_grad = ( - k_grad.sum(dim=0).sum(dim=0) if k_proj_bias is not None else None - ) - v_proj_weight_grad = v_grad.transpose(-1, -2).matmul(context).sum(dim=0) - v_proj_bias_grad = ( - v_grad.sum(dim=0).sum(dim=0) if v_proj_bias is not None else None - ) - - # split grads - q_grad = q_grad.chunk(ctx.sp_size, dim=ctx.dim)[ctx.sp_rank].clone() - k_grad = k_grad.chunk(ctx.sp_size, dim=ctx.dim)[ctx.sp_rank].clone() - v_grad = v_grad.chunk(ctx.sp_size, dim=ctx.dim)[ctx.sp_rank].clone() - - if ctx.is_cross_attn: - hidden_states_grad = torch.matmul(q_grad, q_proj_weight) - context_grad = torch.matmul(k_grad, k_proj_weight) + torch.matmul( - v_grad, v_proj_weight - ) - else: - hidden_states_grad = ( - torch.matmul(q_grad, q_proj_weight) - + torch.matmul(k_grad, k_proj_weight) - + torch.matmul(v_grad, v_proj_weight) - ) - context_grad = hidden_states_grad - - return ( - hidden_states_grad, - context_grad, - q_proj_weight_grad, - q_proj_bias_grad, - k_proj_weight_grad, - k_proj_bias_grad, - v_proj_weight_grad, - v_proj_bias_grad, - None, - None, - None, - None, - ) - - -def async_all_gather_proj_for_two( - hidden_states: torch.Tensor, - context: torch.Tensor, - q_proj_weight: torch.Tensor, - q_proj_bias: Optional[torch.Tensor], - k_proj_weight: torch.Tensor, - k_proj_bias: Optional[torch.Tensor], - v_proj_weight: torch.Tensor, - v_proj_bias: Optional[torch.Tensor], - dim: int, - process_group: dist.ProcessGroup, - sp_size: int, - sp_rank: int, -) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: - return AsyncAllGatherProjForTwo.apply( - hidden_states, - context, - q_proj_weight, - q_proj_bias, - k_proj_weight, - k_proj_bias, - v_proj_weight, - v_proj_bias, - dim, - process_group, - sp_size, - sp_rank, - ) diff --git a/open_sora/utils/data.py b/open_sora/utils/data.py deleted file mode 100644 index a68b760b..00000000 --- a/open_sora/utils/data.py +++ /dev/null @@ -1,356 +0,0 @@ -import os -from typing import List, Optional, Tuple, Union - -import torch -import torch.nn as nn -import torch.nn.functional as F -from colossalai.utils import get_current_device -from datasets import Dataset as HFDataset -from datasets import dataset_dict, load_from_disk -from diffusers.models import AutoencoderKL -from torch.utils.data import ConcatDataset, Dataset -from torchvision.io import read_video -from transformers import AutoModel - -DatasetType = Union[Dataset, ConcatDataset, dataset_dict.Dataset] -PathType = Union[str, os.PathLike] - - -def ceil_to_multiple(x: int, multiple: int) -> int: - m = x % multiple - if m == 0: - return x - return x + multiple - m - - -def video2col(video_4d: torch.Tensor, patch_size: int) -> torch.Tensor: - """ - Convert a 4D video tensor to a 2D tensor where each row is a patch of the video. - - Args: - video_4d (torch.Tensor): A tensor of shape [T, C, H, W] - patch_size (int): The size of the patches. - - Returns: - torch.Tensor: A tensor of shape [S, C, P, P] where S is the number of patches and P is the patch size. - """ - t, c, h, w = video_4d.shape - out = [] - for y in range(0, h, patch_size): - for x in range(0, w, patch_size): - if y + patch_size > h or x + patch_size > w: - continue - patch = video_4d[:, :, y : y + patch_size, x : x + patch_size] - out.append(patch) - # [S, C, P, P] - return torch.stack(out, dim=1).view(-1, c, patch_size, patch_size) - - -def col2video( - patches: torch.Tensor, video_shape: Tuple[int, int, int, int] -) -> torch.Tensor: - """ - Convert a 2D tensor of patches to a 4D video tensor. - - Args: - patches (torch.Tensor): A tensor of shape [S, C, P, P] where S is the number of patches and P is the patch size. - video_shape (Tuple[int, int, int, int]): The shape of the video tensor [T, C, H, W]. - - Returns: - torch.Tensor: A tensor of shape [T, C, H, W]. - """ - t, c, h, w = video_shape - video = torch.empty(t, c, h, w, dtype=patches.dtype, device=patches.device) - patch_size = patches.shape[2] - num_x_patches = w // patch_size - patches = patches.view(t, -1, c, patch_size, patch_size) - for y in range(0, h, patch_size): - for x in range(0, w, patch_size): - if y + patch_size > h or x + patch_size > w: - continue - # [T, C, P, P] - patch = patches[:, (y // patch_size) * num_x_patches + x // patch_size] - video[:, :, y : y + patch_size, x : x + patch_size].copy_(patch) - return video - - -def pad_sequences( - sequences: List[torch.Tensor], pad_to_multiple: Optional[int] = None -) -> Tuple[torch.Tensor, torch.Tensor]: - """Pad a list of sequences. - - Args: - sequences (List[torch.Tensor]): Each sequence is a tensor of shape [T, ...]. - - Returns: - Tuple[torch.Tensor, torch.Tensor]: Padded batch of sequences ([B, T, ...]) and padding mask ([B, T]). - """ - max_len = max([sequence.shape[0] for sequence in sequences]) - if pad_to_multiple is not None: - max_len = ceil_to_multiple(max_len, pad_to_multiple) - padded_sequences = [ - F.pad( - sequence, [0] * (sequence.ndim - 1) * 2 + [0, max_len - sequence.shape[0]] - ) - for sequence in sequences - ] - padded_sequences = torch.stack(padded_sequences, dim=0) - padding_mask = torch.zeros( - padded_sequences.shape[0], - padded_sequences.shape[1], - dtype=torch.int, - device=padded_sequences.device, - ) - for i, sequence in enumerate(sequences): - padding_mask[i, : sequence.shape[0]] = 1 - return padded_sequences, padding_mask - - -def patchify_batch( - videos: List[torch.Tensor], patch_size: int, pad_to_multiple: Optional[int] = None -) -> Tuple[torch.Tensor, torch.Tensor]: - """Patchify a batch of videos. - - Args: - videos (List[torch.Tensor]): A list of tensors of shape [T, C, H, W] - patch_size (int): The size of the patches. - - Returns: - Tuple[torch.Tensor, torch.Tensor]: Padded batch of patches ([B, S, C, P, P]) and padding mask ([B, S]). - """ - video_patches = [video2col(video, patch_size) for video in videos] - return pad_sequences(video_patches, pad_to_multiple=pad_to_multiple) - - -def expand_mask_4d(q_mask: torch.Tensor, kv_mask: torch.Tensor) -> torch.Tensor: - """expand 2D mask to 4D mask - - Args: - q_mask (torch.Tensor): [B, Sq] - kv_mask (torch.Tensor): [B, Skv] - - Returns: - torch.Tensor: [B, 1, Sq, Skv] - """ - q_len = q_mask.shape[1] - mask = kv_mask.unsqueeze(1).repeat(1, q_len, 1) - mask = mask * q_mask.unsqueeze(-1) - return mask.unsqueeze(1) - - -def make_batch( - samples: List[dict], - video_dir: str, - pad_to_multiple: Optional[int] = None, - use_pooled_text: bool = False, -) -> dict: - """Make a batch of samples. - - Args: - samples (List[dict]): A list of samples. - - Returns: - dict: A batch of samples. - """ - videos = [ - read_video(os.path.join(video_dir, sample["video_file"]), pts_unit="sec")[0] - for sample in samples - ] - texts = [sample["text_latent_states"] for sample in samples] - if use_pooled_text: - texts = torch.stack(texts, dim=0) - return { - "videos": videos, - "text_latent_states": texts, - } - texts, text_padding_mask = pad_sequences(texts, pad_to_multiple=pad_to_multiple) - return { - "videos": videos, - "text_latent_states": texts, - "text_padding_mask": text_padding_mask, - } - - -def normalize_video(video: torch.Tensor) -> torch.Tensor: - return video.float() / 255 - 0.5 - - -def unnormalize_video(video: torch.Tensor) -> torch.Tensor: - return (video + 0.5) * 255 - - -class VideoCompressor: - t_factor: int - h_w_factor: int - out_channels: int - - def encode(self, video: torch.Tensor) -> torch.Tensor: - """Encode a video. - - Args: - video (torch.Tensor): [T, H, W, C] - - Returns: - torch.Tensor: [T, C, H, W] - """ - raise NotImplementedError - - def decode(self, latent: torch.Tensor) -> torch.Tensor: - """Decode a latent tensor. - - Args: - latent (torch.Tensor): [T, C, H, W] - - Returns: - torch.Tensor: [T, H, W, C] - """ - raise NotImplementedError - - -class RawVideoCompressor(VideoCompressor): - t_factor = 1 - h_w_factor = 1 - out_channels = 3 - - def encode(self, video: torch.Tensor) -> torch.Tensor: - # [T, H, W, C] -> [T, C, H, W] - return video.permute(0, 3, 1, 2).contiguous() - - def decode(self, latent: torch.Tensor) -> torch.Tensor: - # [T, C, H, W] -> [T, H, W, C] - return latent.permute(0, 2, 3, 1).contiguous() - - -class VqvaeVideoCompressor(VideoCompressor): - t_factor = 2 - h_w_factor = 4 - - def __init__(self, vqvae: nn.Module): - self.vqvae = vqvae - self.out_channels = vqvae.embedding_dim - - def encode(self, video: torch.Tensor) -> torch.Tensor: - # [T, H, W, C] -> [B, C, T, H, W] - video = video.permute(3, 0, 1, 2).unsqueeze(0) - latent_indices, embeddings = self.vqvae.encode(video, include_embeddings=True) - # [B, C, T, H, W] -> [T, C, H, W] - return embeddings.squeeze(0).permute(1, 0, 2, 3) - - def decode(self, latent: torch.Tensor) -> torch.Tensor: - # [T, C, H, W] -> [B, C, T, H, W] - latent = latent.permute(1, 0, 2, 3).unsqueeze(0) - video = self.vqvae.decode_from_embeddings(latent) - # [B, C, T, H, W] -> [T, H, W, C] - video = video.squeeze(0).permute(1, 2, 3, 0) - return video - - -class VaeVideoCompressor(VideoCompressor): - t_factor = 1 - h_w_factor = 8 - out_channels = 4 - - def __init__(self, vae: nn.Module): - self.vae = vae - - def encode(self, video: torch.Tensor) -> torch.Tensor: - # [T, H, W, C] -> [T, C, H, W] - video = video.permute(0, 3, 1, 2) - return self.vae.encode(video).latent_dist.sample().mul_(0.18215) - - def decode(self, latent: torch.Tensor) -> torch.Tensor: - video = self.vae.decode(latent / 0.18215).sample - # [T, C, H, W] -> [T, H, W, C] - return video.permute(0, 2, 3, 1).contiguous() - - -def create_video_compressor( - compressor_type: str, - vqvae_path="hpcai-tech/vqvae", - vae_path="stabilityai/sd-vae-ft-mse", -) -> VideoCompressor: - if compressor_type == "raw": - return RawVideoCompressor() - if compressor_type == "vqvae": - vqvae = ( - AutoModel.from_pretrained(vqvae_path, trust_remote_code=True) - .to(get_current_device()) - .eval() - ) - return VqvaeVideoCompressor(vqvae) - if compressor_type == "vae": - vae = AutoencoderKL.from_pretrained(vae_path).to(get_current_device()).eval() - return VaeVideoCompressor(vae) - raise ValueError(f"Unsupported video compressor type {compressor_type}") - - -@torch.no_grad() -def preprocess_batch( - batch: dict, - patch_size: int, - video_compressor: VideoCompressor, - device=None, - pad_to_multiple: Optional[int] = None, - model_arch: str = "cross-attn", -) -> dict: - if device is None: - device = get_current_device() - videos = [] - for video in batch.pop("videos"): - video = video.to(device) - video = normalize_video(video) - video = video_compressor.encode(video) - videos.append(video) - video_latent_states, video_padding_mask = patchify_batch( - videos, patch_size, pad_to_multiple - ) - batch["video_latent_states"] = video_latent_states - batch["video_padding_mask"] = video_padding_mask - if model_arch == "adaln": - batch["attention_mask"] = expand_mask_4d(video_padding_mask, video_padding_mask) - else: - text_padding_mask = batch.pop("text_padding_mask").to(device) - if model_arch == "cross-attn": - batch["attention_mask"] = expand_mask_4d( - video_padding_mask, text_padding_mask - ) - else: - attention_mask = torch.cat([text_padding_mask, video_padding_mask], dim=1) - batch["attention_mask"] = expand_mask_4d(attention_mask, attention_mask) - batch["text_latent_states"] = batch["text_latent_states"].to(device) - return batch - - -def load_datasets( - dataset_paths: Union[PathType, List[PathType]], mode: str = "train" -) -> Optional[DatasetType]: - """ - Load pre-tokenized dataset. - Each instance of dataset is a dictionary with - `{'input_ids': List[int], 'labels': List[int], sequence: str}` format. - """ - mode_map = {"train": "train", "dev": "validation", "test": "test"} - assert mode in tuple( - mode_map - ), f"Unsupported mode {mode}, it must be in {tuple(mode_map)}" - - if isinstance(dataset_paths, (str, os.PathLike)): - dataset_paths = [dataset_paths] - - datasets = [] # `List[datasets.dataset_dict.Dataset]` - for ds_path in dataset_paths: - ds_path = os.path.abspath(ds_path) - assert os.path.exists(ds_path), f"Not existed file path {ds_path}" - ds_dict = load_from_disk( - dataset_path=ds_path, keep_in_memory=False - ).with_format("torch") - if isinstance(ds_dict, HFDataset): - datasets.append(ds_dict) - else: - if mode_map[mode] in ds_dict: - datasets.append(ds_dict[mode_map[mode]]) - if len(datasets) == 0: - return None - if len(datasets) == 1: - return datasets.pop() - return ConcatDataset(datasets=datasets) diff --git a/opensora/__init__.py b/opensora/__init__.py new file mode 100644 index 00000000..a3175b2d --- /dev/null +++ b/opensora/__init__.py @@ -0,0 +1,4 @@ +from .acceleration import * +from .datasets import * +from .models import * +from .registry import * diff --git a/open_sora/utils/__init__.py b/opensora/acceleration/__init__.py similarity index 100% rename from open_sora/utils/__init__.py rename to opensora/acceleration/__init__.py diff --git a/opensora/acceleration/checkpoint.py b/opensora/acceleration/checkpoint.py new file mode 100644 index 00000000..d832a010 --- /dev/null +++ b/opensora/acceleration/checkpoint.py @@ -0,0 +1,24 @@ +from collections.abc import Iterable + +import torch.nn as nn +from torch.utils.checkpoint import checkpoint, checkpoint_sequential + + +def set_grad_checkpoint(model, use_fp32_attention=False, gc_step=1): + assert isinstance(model, nn.Module) + + def set_attr(module): + module.grad_checkpointing = True + module.fp32_attention = use_fp32_attention + module.grad_checkpointing_step = gc_step + + model.apply(set_attr) + + +def auto_grad_checkpoint(module, *args, **kwargs): + if getattr(module, "grad_checkpointing", False): + if not isinstance(module, Iterable): + return checkpoint(module, *args, **kwargs) + gc_step = module[0].grad_checkpointing_step + return checkpoint_sequential(module, gc_step, *args, **kwargs) + return module(*args, **kwargs) diff --git a/opensora/acceleration/communications.py b/opensora/acceleration/communications.py new file mode 100644 index 00000000..d0900d20 --- /dev/null +++ b/opensora/acceleration/communications.py @@ -0,0 +1,188 @@ +import torch +import torch.distributed as dist + + +# ==================== +# All-To-All +# ==================== +def _all_to_all( + input_: torch.Tensor, + world_size: int, + group: dist.ProcessGroup, + scatter_dim: int, + gather_dim: int, +): + input_list = [t.contiguous() for t in torch.tensor_split(input_, world_size, scatter_dim)] + output_list = [torch.empty_like(input_list[0]) for _ in range(world_size)] + dist.all_to_all(output_list, input_list, group=group) + return torch.cat(output_list, dim=gather_dim).contiguous() + + +class _AllToAll(torch.autograd.Function): + """All-to-all communication. + + Args: + input_: input matrix + process_group: communication group + scatter_dim: scatter dimension + gather_dim: gather dimension + """ + + @staticmethod + def forward(ctx, input_, process_group, scatter_dim, gather_dim): + ctx.process_group = process_group + ctx.scatter_dim = scatter_dim + ctx.gather_dim = gather_dim + ctx.world_size = dist.get_world_size(process_group) + output = _all_to_all(input_, ctx.world_size, process_group, scatter_dim, gather_dim) + return output + + @staticmethod + def backward(ctx, grad_output): + grad_output = _all_to_all( + grad_output, + ctx.world_size, + ctx.process_group, + ctx.gather_dim, + ctx.scatter_dim, + ) + return ( + grad_output, + None, + None, + None, + ) + + +def all_to_all( + input_: torch.Tensor, + process_group: dist.ProcessGroup, + scatter_dim: int = 2, + gather_dim: int = 1, +): + return _AllToAll.apply(input_, process_group, scatter_dim, gather_dim) + + +def _gather( + input_: torch.Tensor, + world_size: int, + group: dist.ProcessGroup, + gather_dim: int, +): + if gather_list is None: + gather_list = [torch.empty_like(input_) for _ in range(world_size)] + dist.gather(input_, gather_list, group=group, gather_dim=gather_dim) + return gather_list + + +# ==================== +# Gather-Split +# ==================== + + +def _split(input_, pg: dist.ProcessGroup, dim=-1): + # skip if only one rank involved + world_size = dist.get_world_size(pg) + rank = dist.get_rank(pg) + if world_size == 1: + return input_ + + # Split along last dimension. + dim_size = input_.size(dim) + assert dim_size % world_size == 0, ( + f"The dimension to split ({dim_size}) is not a multiple of world size ({world_size}), " + f"cannot split tensor evenly" + ) + + tensor_list = torch.split(input_, dim_size // world_size, dim=dim) + output = tensor_list[rank].contiguous() + + return output + + +def _gather(input_, pg: dist.ProcessGroup, dim=-1): + # skip if only one rank involved + input_ = input_.contiguous() + world_size = dist.get_world_size(pg) + dist.get_rank(pg) + + if world_size == 1: + return input_ + + # all gather + tensor_list = [torch.empty_like(input_) for _ in range(world_size)] + assert input_.device.type == "cuda" + torch.distributed.all_gather(tensor_list, input_, group=pg) + + # concat + output = torch.cat(tensor_list, dim=dim).contiguous() + + return output + + +class _GatherForwardSplitBackward(torch.autograd.Function): + """Gather the input from model parallel region and concatenate. + + Args: + input_: input matrix. + process_group: parallel mode. + dim: dimension + """ + + @staticmethod + def symbolic(graph, input_): + return _gather(input_) + + @staticmethod + def forward(ctx, input_, process_group, dim, grad_scale): + ctx.mode = process_group + ctx.dim = dim + ctx.grad_scale = grad_scale + return _gather(input_, process_group, dim) + + @staticmethod + def backward(ctx, grad_output): + if ctx.grad_scale == "up": + grad_output = grad_output * dist.get_world_size(ctx.mode) + elif ctx.grad_scale == "down": + grad_output = grad_output / dist.get_world_size(ctx.mode) + + return _split(grad_output, ctx.mode, ctx.dim), None, None, None + + +class _SplitForwardGatherBackward(torch.autograd.Function): + """ + Split the input and keep only the corresponding chuck to the rank. + + Args: + input_: input matrix. + process_group: parallel mode. + dim: dimension + """ + + @staticmethod + def symbolic(graph, input_): + return _split(input_) + + @staticmethod + def forward(ctx, input_, process_group, dim, grad_scale): + ctx.mode = process_group + ctx.dim = dim + ctx.grad_scale = grad_scale + return _split(input_, process_group, dim) + + @staticmethod + def backward(ctx, grad_output): + if ctx.grad_scale == "up": + grad_output = grad_output * dist.get_world_size(ctx.mode) + elif ctx.grad_scale == "down": + grad_output = grad_output / dist.get_world_size(ctx.mode) + return _gather(grad_output, ctx.mode, ctx.dim), None, None, None + + +def split_forward_gather_backward(input_, process_group, dim, grad_scale=1.0): + return _SplitForwardGatherBackward.apply(input_, process_group, dim, grad_scale) + + +def gather_forward_split_backward(input_, process_group, dim, grad_scale=None): + return _GatherForwardSplitBackward.apply(input_, process_group, dim, grad_scale) diff --git a/opensora/acceleration/parallel_states.py b/opensora/acceleration/parallel_states.py new file mode 100644 index 00000000..ff2893e3 --- /dev/null +++ b/opensora/acceleration/parallel_states.py @@ -0,0 +1,19 @@ +import torch.distributed as dist + +_GLOBAL_PARALLEL_GROUPS = dict() + + +def set_data_parallel_group(group: dist.ProcessGroup): + _GLOBAL_PARALLEL_GROUPS["data"] = group + + +def get_data_parallel_group(): + return _GLOBAL_PARALLEL_GROUPS.get("data", None) + + +def set_sequence_parallel_group(group: dist.ProcessGroup): + _GLOBAL_PARALLEL_GROUPS["sequence"] = group + + +def get_sequence_parallel_group(): + return _GLOBAL_PARALLEL_GROUPS.get("sequence", None) diff --git a/open_sora/utils/plugin.py b/opensora/acceleration/plugin.py similarity index 95% rename from open_sora/utils/plugin.py rename to opensora/acceleration/plugin.py index f687b413..c657a953 100644 --- a/open_sora/utils/plugin.py +++ b/opensora/acceleration/plugin.py @@ -75,13 +75,11 @@ def prepare_dataloader( pin_memory=False, num_workers=0, distributed_sampler_cls=None, - **kwargs + **kwargs, ): _kwargs = kwargs.copy() distributed_sampler_cls = distributed_sampler_cls or DistributedSampler - sampler = distributed_sampler_cls( - dataset, num_replicas=self.dp_size, rank=self.dp_rank, shuffle=shuffle - ) + sampler = distributed_sampler_cls(dataset, num_replicas=self.dp_size, rank=self.dp_rank, shuffle=shuffle) # Deterministic dataloader def seed_worker(worker_id): diff --git a/opensora/acceleration/shardformer/modeling/__init__.py b/opensora/acceleration/shardformer/modeling/__init__.py new file mode 100644 index 00000000..e69de29b diff --git a/opensora/acceleration/shardformer/modeling/t5.py b/opensora/acceleration/shardformer/modeling/t5.py new file mode 100644 index 00000000..9cfb8084 --- /dev/null +++ b/opensora/acceleration/shardformer/modeling/t5.py @@ -0,0 +1,39 @@ +import torch +import torch.nn as nn + + +class T5LayerNorm(nn.Module): + def __init__(self, hidden_size, eps=1e-6): + """ + Construct a layernorm module in the T5 style. No bias and no subtraction of mean. + """ + super().__init__() + self.weight = nn.Parameter(torch.ones(hidden_size)) + self.variance_epsilon = eps + + def forward(self, hidden_states): + # T5 uses a layer_norm which only scales and doesn't shift, which is also known as Root Mean + # Square Layer Normalization https://arxiv.org/abs/1910.07467 thus varience is calculated + # w/o mean and there is no bias. Additionally we want to make sure that the accumulation for + # half-precision inputs is done in fp32 + + variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True) + hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) + + # convert into half-precision if necessary + if self.weight.dtype in [torch.float16, torch.bfloat16]: + hidden_states = hidden_states.to(self.weight.dtype) + + return self.weight * hidden_states + + @staticmethod + def from_native_module(module, *args, **kwargs): + assert module.__class__.__name__ == "FusedRMSNorm", ( + "Recovering T5LayerNorm requires the original layer to be apex's Fused RMS Norm." + "Apex's fused norm is automatically used by Hugging Face Transformers https://github.com/huggingface/transformers/blob/main/src/transformers/models/t5/modeling_t5.py#L265C5-L265C48" + ) + + layer_norm = T5LayerNorm(module.normalized_shape, eps=module.eps) + layer_norm.weight.data.copy_(module.weight.data) + layer_norm = layer_norm.to(module.weight.device) + return layer_norm diff --git a/opensora/acceleration/shardformer/policy/__init__.py b/opensora/acceleration/shardformer/policy/__init__.py new file mode 100644 index 00000000..e69de29b diff --git a/opensora/acceleration/shardformer/policy/t5_encoder.py b/opensora/acceleration/shardformer/policy/t5_encoder.py new file mode 100644 index 00000000..85c994ec --- /dev/null +++ b/opensora/acceleration/shardformer/policy/t5_encoder.py @@ -0,0 +1,67 @@ +from colossalai.shardformer.modeling.jit import get_jit_fused_dropout_add_func +from colossalai.shardformer.modeling.t5 import get_jit_fused_T5_layer_ff_forward, get_T5_layer_self_attention_forward +from colossalai.shardformer.policies.base_policy import Policy, SubModuleReplacementDescription + + +class T5EncoderPolicy(Policy): + def config_sanity_check(self): + assert not self.shard_config.enable_tensor_parallelism + assert not self.shard_config.enable_flash_attention + + def preprocess(self): + return self.model + + def module_policy(self): + from transformers.models.t5.modeling_t5 import T5LayerFF, T5LayerSelfAttention, T5Stack + + policy = {} + + # check whether apex is installed + try: + from opensora.acceleration.shardformer.modeling.t5 import T5LayerNorm + + # recover hf from fused rms norm to T5 norm which is faster + self.append_or_create_submodule_replacement( + description=SubModuleReplacementDescription( + suffix="layer_norm", + target_module=T5LayerNorm, + ), + policy=policy, + target_key=T5LayerFF, + ) + self.append_or_create_submodule_replacement( + description=SubModuleReplacementDescription(suffix="layer_norm", target_module=T5LayerNorm), + policy=policy, + target_key=T5LayerSelfAttention, + ) + self.append_or_create_submodule_replacement( + description=SubModuleReplacementDescription(suffix="final_layer_norm", target_module=T5LayerNorm), + policy=policy, + target_key=T5Stack, + ) + except (ImportError, ModuleNotFoundError): + pass + + # use jit operator + if self.shard_config.enable_jit_fused: + self.append_or_create_method_replacement( + description={ + "forward": get_jit_fused_T5_layer_ff_forward(), + "dropout_add": get_jit_fused_dropout_add_func(), + }, + policy=policy, + target_key=T5LayerFF, + ) + self.append_or_create_method_replacement( + description={ + "forward": get_T5_layer_self_attention_forward(), + "dropout_add": get_jit_fused_dropout_add_func(), + }, + policy=policy, + target_key=T5LayerSelfAttention, + ) + + return policy + + def postprocess(self): + return self.model diff --git a/opensora/datasets/__init__.py b/opensora/datasets/__init__.py new file mode 100644 index 00000000..c9b33954 --- /dev/null +++ b/opensora/datasets/__init__.py @@ -0,0 +1,2 @@ +from .datasets import DatasetFromCSV, get_transforms_image, get_transforms_video +from .utils import prepare_dataloader, save_sample diff --git a/opensora/datasets/datasets.py b/opensora/datasets/datasets.py new file mode 100644 index 00000000..9d931727 --- /dev/null +++ b/opensora/datasets/datasets.py @@ -0,0 +1,114 @@ +import csv +import os + +import numpy as np +import torch +import torchvision +import torchvision.transforms as transforms +from torchvision.datasets.folder import IMG_EXTENSIONS, pil_loader + +from . import video_transforms +from .utils import center_crop_arr + + +def get_transforms_video(resolution=256): + transform_video = transforms.Compose( + [ + video_transforms.ToTensorVideo(), # TCHW + video_transforms.RandomHorizontalFlipVideo(), + video_transforms.UCFCenterCropVideo(resolution), + transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True), + ] + ) + return transform_video + + +def get_transforms_image(image_size=256): + transform = transforms.Compose( + [ + transforms.Lambda(lambda pil_image: center_crop_arr(pil_image, image_size)), + transforms.RandomHorizontalFlip(), + transforms.ToTensor(), + transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True), + ] + ) + return transform + + +class DatasetFromCSV(torch.utils.data.Dataset): + """load video according to the csv file. + + Args: + target_video_len (int): the number of video frames will be load. + align_transform (callable): Align different videos in a specified size. + temporal_sample (callable): Sample the target length of a video. + """ + + def __init__( + self, + csv_path, + num_frames=16, + frame_interval=1, + transform=None, + root=None, + ): + self.csv_path = csv_path + with open(csv_path, "r") as f: + reader = csv.reader(f) + self.samples = list(reader) + + ext = self.samples[0][0].split(".")[-1] + if ext.lower() in ("mp4", "avi", "mov", "mkv"): + self.is_video = True + else: + assert f".{ext.lower()}" in IMG_EXTENSIONS, f"Unsupported file format: {ext}" + self.is_video = False + + self.transform = transform + + self.num_frames = num_frames + self.frame_interval = frame_interval + self.temporal_sample = video_transforms.TemporalRandomCrop(num_frames * frame_interval) + self.root = root + + def getitem(self, index): + sample = self.samples[index] + path = sample[0] + if self.root: + path = os.path.join(self.root, path) + text = sample[1] + + if self.is_video: + vframes, aframes, info = torchvision.io.read_video(filename=path, pts_unit="sec", output_format="TCHW") + total_frames = len(vframes) + + # Sampling video frames + start_frame_ind, end_frame_ind = self.temporal_sample(total_frames) + assert ( + end_frame_ind - start_frame_ind >= self.num_frames + ), f"{path} with index {index} has not enough frames." + frame_indice = np.linspace(start_frame_ind, end_frame_ind - 1, self.num_frames, dtype=int) + + video = vframes[frame_indice] + video = self.transform(video) # T C H W + else: + image = pil_loader(path) + image = self.transform(image) + video = image.unsqueeze(0).repeat(self.num_frames, 1, 1, 1) + + # TCHW -> CTHW + video = video.permute(1, 0, 2, 3) + + return {"video": video, "text": text} + + def __getitem__(self, index): + for _ in range(10): + try: + return self.getitem(index) + except Exception as e: + print(e) + index = np.random.randint(len(self)) + raise RuntimeError("Too many bad data.") + + def __len__(self): + return len(self.samples) diff --git a/opensora/datasets/utils.py b/opensora/datasets/utils.py new file mode 100644 index 00000000..cd268ae6 --- /dev/null +++ b/opensora/datasets/utils.py @@ -0,0 +1,135 @@ +import random +from typing import Iterator, Optional + +import numpy as np +import torch +from PIL import Image +from torch.distributed import ProcessGroup +from torch.distributed.distributed_c10d import _get_default_group +from torch.utils.data import DataLoader, Dataset +from torch.utils.data.distributed import DistributedSampler +from torchvision.io import write_video +from torchvision.utils import save_image + + +def save_sample(x, fps=8, save_path=None, normalize=True, value_range=(-1, 1)): + """ + Args: + x (Tensor): shape [C, T, H, W] + """ + assert x.ndim == 4 + + if x.shape[1] == 1: # T = 1: save as image + save_path += ".png" + x = x.squeeze(1) + save_image([x], save_path, normalize=normalize, value_range=value_range) + else: + save_path += ".mp4" + if normalize: + low, high = value_range + x.clamp_(min=low, max=high) + x.sub_(low).div_(max(high - low, 1e-5)) + + x = x.mul(255).add_(0.5).clamp_(0, 255).permute(1, 2, 3, 0).to("cpu", torch.uint8) + write_video(save_path, x, fps=fps, video_codec="h264") + print(f"Saved to {save_path}") + + +class StatefulDistributedSampler(DistributedSampler): + def __init__( + self, + dataset: Dataset, + num_replicas: Optional[int] = None, + rank: Optional[int] = None, + shuffle: bool = True, + seed: int = 0, + drop_last: bool = False, + ) -> None: + super().__init__(dataset, num_replicas, rank, shuffle, seed, drop_last) + self.start_index: int = 0 + + def __iter__(self) -> Iterator: + iterator = super().__iter__() + indices = list(iterator) + indices = indices[self.start_index :] + return iter(indices) + + def __len__(self) -> int: + return self.num_samples - self.start_index + + def set_start_index(self, start_index: int) -> None: + self.start_index = start_index + + +def prepare_dataloader( + dataset, + batch_size, + shuffle=False, + seed=1024, + drop_last=False, + pin_memory=False, + num_workers=0, + process_group: Optional[ProcessGroup] = None, + **kwargs, +): + r""" + Prepare a dataloader for distributed training. The dataloader will be wrapped by + `torch.utils.data.DataLoader` and `StatefulDistributedSampler`. + + + Args: + dataset (`torch.utils.data.Dataset`): The dataset to be loaded. + shuffle (bool, optional): Whether to shuffle the dataset. Defaults to False. + seed (int, optional): Random worker seed for sampling, defaults to 1024. + add_sampler: Whether to add ``DistributedDataParallelSampler`` to the dataset. Defaults to True. + drop_last (bool, optional): Set to True to drop the last incomplete batch, if the dataset size + is not divisible by the batch size. If False and the size of dataset is not divisible by + the batch size, then the last batch will be smaller, defaults to False. + pin_memory (bool, optional): Whether to pin memory address in CPU memory. Defaults to False. + num_workers (int, optional): Number of worker threads for this dataloader. Defaults to 0. + kwargs (dict): optional parameters for ``torch.utils.data.DataLoader``, more details could be found in + `DataLoader `_. + + Returns: + :class:`torch.utils.data.DataLoader`: A DataLoader used for training or testing. + """ + _kwargs = kwargs.copy() + process_group = process_group or _get_default_group() + sampler = StatefulDistributedSampler( + dataset, num_replicas=process_group.size(), rank=process_group.rank(), shuffle=shuffle + ) + + # Deterministic dataloader + def seed_worker(worker_id): + worker_seed = seed + np.random.seed(worker_seed) + torch.manual_seed(worker_seed) + random.seed(worker_seed) + + return DataLoader( + dataset, + batch_size=batch_size, + sampler=sampler, + worker_init_fn=seed_worker, + drop_last=drop_last, + pin_memory=pin_memory, + num_workers=num_workers, + **_kwargs, + ) + + +def center_crop_arr(pil_image, image_size): + """ + Center cropping implementation from ADM. + https://github.com/openai/guided-diffusion/blob/8fb3ad9197f16bbc40620447b2742e13458d2831/guided_diffusion/image_datasets.py#L126 + """ + while min(*pil_image.size) >= 2 * image_size: + pil_image = pil_image.resize(tuple(x // 2 for x in pil_image.size), resample=Image.BOX) + + scale = image_size / min(*pil_image.size) + pil_image = pil_image.resize(tuple(round(x * scale) for x in pil_image.size), resample=Image.BICUBIC) + + arr = np.array(pil_image) + crop_y = (arr.shape[0] - image_size) // 2 + crop_x = (arr.shape[1] - image_size) // 2 + return Image.fromarray(arr[crop_y : crop_y + image_size, crop_x : crop_x + image_size]) diff --git a/opensora/datasets/video_transforms.py b/opensora/datasets/video_transforms.py new file mode 100644 index 00000000..a0d1cec8 --- /dev/null +++ b/opensora/datasets/video_transforms.py @@ -0,0 +1,501 @@ +# Copyright 2024 Vchitect/Latte + +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at + +# http://www.apache.org/licenses/LICENSE-2.0 + +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License.# Modified from Latte + +# - This file is adapted from https://github.com/Vchitect/Latte/blob/main/datasets/video_transforms.py + + +import numbers +import random + +import torch + + +def _is_tensor_video_clip(clip): + if not torch.is_tensor(clip): + raise TypeError("clip should be Tensor. Got %s" % type(clip)) + + if not clip.ndimension() == 4: + raise ValueError("clip should be 4D. Got %dD" % clip.dim()) + + return True + + +def center_crop_arr(pil_image, image_size): + """ + Center cropping implementation from ADM. + https://github.com/openai/guided-diffusion/blob/8fb3ad9197f16bbc40620447b2742e13458d2831/guided_diffusion/image_datasets.py#L126 + """ + while min(*pil_image.size) >= 2 * image_size: + pil_image = pil_image.resize(tuple(x // 2 for x in pil_image.size), resample=Image.BOX) + + scale = image_size / min(*pil_image.size) + pil_image = pil_image.resize(tuple(round(x * scale) for x in pil_image.size), resample=Image.BICUBIC) + + arr = np.array(pil_image) + crop_y = (arr.shape[0] - image_size) // 2 + crop_x = (arr.shape[1] - image_size) // 2 + return Image.fromarray(arr[crop_y : crop_y + image_size, crop_x : crop_x + image_size]) + + +def crop(clip, i, j, h, w): + """ + Args: + clip (torch.tensor): Video clip to be cropped. Size is (T, C, H, W) + """ + if len(clip.size()) != 4: + raise ValueError("clip should be a 4D tensor") + return clip[..., i : i + h, j : j + w] + + +def resize(clip, target_size, interpolation_mode): + if len(target_size) != 2: + raise ValueError(f"target size should be tuple (height, width), instead got {target_size}") + return torch.nn.functional.interpolate(clip, size=target_size, mode=interpolation_mode, align_corners=False) + + +def resize_scale(clip, target_size, interpolation_mode): + if len(target_size) != 2: + raise ValueError(f"target size should be tuple (height, width), instead got {target_size}") + H, W = clip.size(-2), clip.size(-1) + scale_ = target_size[0] / min(H, W) + return torch.nn.functional.interpolate(clip, scale_factor=scale_, mode=interpolation_mode, align_corners=False) + + +def resized_crop(clip, i, j, h, w, size, interpolation_mode="bilinear"): + """ + Do spatial cropping and resizing to the video clip + Args: + clip (torch.tensor): Video clip to be cropped. Size is (T, C, H, W) + i (int): i in (i,j) i.e coordinates of the upper left corner. + j (int): j in (i,j) i.e coordinates of the upper left corner. + h (int): Height of the cropped region. + w (int): Width of the cropped region. + size (tuple(int, int)): height and width of resized clip + Returns: + clip (torch.tensor): Resized and cropped clip. Size is (T, C, H, W) + """ + if not _is_tensor_video_clip(clip): + raise ValueError("clip should be a 4D torch.tensor") + clip = crop(clip, i, j, h, w) + clip = resize(clip, size, interpolation_mode) + return clip + + +def center_crop(clip, crop_size): + if not _is_tensor_video_clip(clip): + raise ValueError("clip should be a 4D torch.tensor") + h, w = clip.size(-2), clip.size(-1) + th, tw = crop_size + if h < th or w < tw: + raise ValueError("height and width must be no smaller than crop_size") + + i = int(round((h - th) / 2.0)) + j = int(round((w - tw) / 2.0)) + return crop(clip, i, j, th, tw) + + +def center_crop_using_short_edge(clip): + if not _is_tensor_video_clip(clip): + raise ValueError("clip should be a 4D torch.tensor") + h, w = clip.size(-2), clip.size(-1) + if h < w: + th, tw = h, h + i = 0 + j = int(round((w - tw) / 2.0)) + else: + th, tw = w, w + i = int(round((h - th) / 2.0)) + j = 0 + return crop(clip, i, j, th, tw) + + +def random_shift_crop(clip): + """ + Slide along the long edge, with the short edge as crop size + """ + if not _is_tensor_video_clip(clip): + raise ValueError("clip should be a 4D torch.tensor") + h, w = clip.size(-2), clip.size(-1) + + if h <= w: + short_edge = h + else: + short_edge = w + + th, tw = short_edge, short_edge + + i = torch.randint(0, h - th + 1, size=(1,)).item() + j = torch.randint(0, w - tw + 1, size=(1,)).item() + return crop(clip, i, j, th, tw) + + +def to_tensor(clip): + """ + Convert tensor data type from uint8 to float, divide value by 255.0 and + permute the dimensions of clip tensor + Args: + clip (torch.tensor, dtype=torch.uint8): Size is (T, C, H, W) + Return: + clip (torch.tensor, dtype=torch.float): Size is (T, C, H, W) + """ + _is_tensor_video_clip(clip) + if not clip.dtype == torch.uint8: + raise TypeError("clip tensor should have data type uint8. Got %s" % str(clip.dtype)) + # return clip.float().permute(3, 0, 1, 2) / 255.0 + return clip.float() / 255.0 + + +def normalize(clip, mean, std, inplace=False): + """ + Args: + clip (torch.tensor): Video clip to be normalized. Size is (T, C, H, W) + mean (tuple): pixel RGB mean. Size is (3) + std (tuple): pixel standard deviation. Size is (3) + Returns: + normalized clip (torch.tensor): Size is (T, C, H, W) + """ + if not _is_tensor_video_clip(clip): + raise ValueError("clip should be a 4D torch.tensor") + if not inplace: + clip = clip.clone() + mean = torch.as_tensor(mean, dtype=clip.dtype, device=clip.device) + # print(mean) + std = torch.as_tensor(std, dtype=clip.dtype, device=clip.device) + clip.sub_(mean[:, None, None, None]).div_(std[:, None, None, None]) + return clip + + +def hflip(clip): + """ + Args: + clip (torch.tensor): Video clip to be normalized. Size is (T, C, H, W) + Returns: + flipped clip (torch.tensor): Size is (T, C, H, W) + """ + if not _is_tensor_video_clip(clip): + raise ValueError("clip should be a 4D torch.tensor") + return clip.flip(-1) + + +class RandomCropVideo: + def __init__(self, size): + if isinstance(size, numbers.Number): + self.size = (int(size), int(size)) + else: + self.size = size + + def __call__(self, clip): + """ + Args: + clip (torch.tensor): Video clip to be cropped. Size is (T, C, H, W) + Returns: + torch.tensor: randomly cropped video clip. + size is (T, C, OH, OW) + """ + i, j, h, w = self.get_params(clip) + return crop(clip, i, j, h, w) + + def get_params(self, clip): + h, w = clip.shape[-2:] + th, tw = self.size + + if h < th or w < tw: + raise ValueError(f"Required crop size {(th, tw)} is larger than input image size {(h, w)}") + + if w == tw and h == th: + return 0, 0, h, w + + i = torch.randint(0, h - th + 1, size=(1,)).item() + j = torch.randint(0, w - tw + 1, size=(1,)).item() + + return i, j, th, tw + + def __repr__(self) -> str: + return f"{self.__class__.__name__}(size={self.size})" + + +class CenterCropResizeVideo: + """ + First use the short side for cropping length, + center crop video, then resize to the specified size + """ + + def __init__( + self, + size, + interpolation_mode="bilinear", + ): + if isinstance(size, tuple): + if len(size) != 2: + raise ValueError(f"size should be tuple (height, width), instead got {size}") + self.size = size + else: + self.size = (size, size) + + self.interpolation_mode = interpolation_mode + + def __call__(self, clip): + """ + Args: + clip (torch.tensor): Video clip to be cropped. Size is (T, C, H, W) + Returns: + torch.tensor: scale resized / center cropped video clip. + size is (T, C, crop_size, crop_size) + """ + clip_center_crop = center_crop_using_short_edge(clip) + clip_center_crop_resize = resize( + clip_center_crop, target_size=self.size, interpolation_mode=self.interpolation_mode + ) + return clip_center_crop_resize + + def __repr__(self) -> str: + return f"{self.__class__.__name__}(size={self.size}, interpolation_mode={self.interpolation_mode}" + + +class UCFCenterCropVideo: + """ + First scale to the specified size in equal proportion to the short edge, + then center cropping + """ + + def __init__( + self, + size, + interpolation_mode="bilinear", + ): + if isinstance(size, tuple): + if len(size) != 2: + raise ValueError(f"size should be tuple (height, width), instead got {size}") + self.size = size + else: + self.size = (size, size) + + self.interpolation_mode = interpolation_mode + + def __call__(self, clip): + """ + Args: + clip (torch.tensor): Video clip to be cropped. Size is (T, C, H, W) + Returns: + torch.tensor: scale resized / center cropped video clip. + size is (T, C, crop_size, crop_size) + """ + clip_resize = resize_scale(clip=clip, target_size=self.size, interpolation_mode=self.interpolation_mode) + clip_center_crop = center_crop(clip_resize, self.size) + return clip_center_crop + + def __repr__(self) -> str: + return f"{self.__class__.__name__}(size={self.size}, interpolation_mode={self.interpolation_mode}" + + +class KineticsRandomCropResizeVideo: + """ + Slide along the long edge, with the short edge as crop size. And resie to the desired size. + """ + + def __init__( + self, + size, + interpolation_mode="bilinear", + ): + if isinstance(size, tuple): + if len(size) != 2: + raise ValueError(f"size should be tuple (height, width), instead got {size}") + self.size = size + else: + self.size = (size, size) + + self.interpolation_mode = interpolation_mode + + def __call__(self, clip): + clip_random_crop = random_shift_crop(clip) + clip_resize = resize(clip_random_crop, self.size, self.interpolation_mode) + return clip_resize + + +class CenterCropVideo: + def __init__( + self, + size, + interpolation_mode="bilinear", + ): + if isinstance(size, tuple): + if len(size) != 2: + raise ValueError(f"size should be tuple (height, width), instead got {size}") + self.size = size + else: + self.size = (size, size) + + self.interpolation_mode = interpolation_mode + + def __call__(self, clip): + """ + Args: + clip (torch.tensor): Video clip to be cropped. Size is (T, C, H, W) + Returns: + torch.tensor: center cropped video clip. + size is (T, C, crop_size, crop_size) + """ + clip_center_crop = center_crop(clip, self.size) + return clip_center_crop + + def __repr__(self) -> str: + return f"{self.__class__.__name__}(size={self.size}, interpolation_mode={self.interpolation_mode}" + + +class NormalizeVideo: + """ + Normalize the video clip by mean subtraction and division by standard deviation + Args: + mean (3-tuple): pixel RGB mean + std (3-tuple): pixel RGB standard deviation + inplace (boolean): whether do in-place normalization + """ + + def __init__(self, mean, std, inplace=False): + self.mean = mean + self.std = std + self.inplace = inplace + + def __call__(self, clip): + """ + Args: + clip (torch.tensor): video clip must be normalized. Size is (C, T, H, W) + """ + return normalize(clip, self.mean, self.std, self.inplace) + + def __repr__(self) -> str: + return f"{self.__class__.__name__}(mean={self.mean}, std={self.std}, inplace={self.inplace})" + + +class ToTensorVideo: + """ + Convert tensor data type from uint8 to float, divide value by 255.0 and + permute the dimensions of clip tensor + """ + + def __init__(self): + pass + + def __call__(self, clip): + """ + Args: + clip (torch.tensor, dtype=torch.uint8): Size is (T, C, H, W) + Return: + clip (torch.tensor, dtype=torch.float): Size is (T, C, H, W) + """ + return to_tensor(clip) + + def __repr__(self) -> str: + return self.__class__.__name__ + + +class RandomHorizontalFlipVideo: + """ + Flip the video clip along the horizontal direction with a given probability + Args: + p (float): probability of the clip being flipped. Default value is 0.5 + """ + + def __init__(self, p=0.5): + self.p = p + + def __call__(self, clip): + """ + Args: + clip (torch.tensor): Size is (T, C, H, W) + Return: + clip (torch.tensor): Size is (T, C, H, W) + """ + if random.random() < self.p: + clip = hflip(clip) + return clip + + def __repr__(self) -> str: + return f"{self.__class__.__name__}(p={self.p})" + + +# ------------------------------------------------------------ +# --------------------- Sampling --------------------------- +# ------------------------------------------------------------ +class TemporalRandomCrop(object): + """Temporally crop the given frame indices at a random location. + + Args: + size (int): Desired length of frames will be seen in the model. + """ + + def __init__(self, size): + self.size = size + + def __call__(self, total_frames): + rand_end = max(0, total_frames - self.size - 1) + begin_index = random.randint(0, rand_end) + end_index = min(begin_index + self.size, total_frames) + return begin_index, end_index + + +if __name__ == "__main__": + import os + + import numpy as np + import torchvision.io as io + from torchvision import transforms + from torchvision.utils import save_image + + vframes, aframes, info = io.read_video(filename="./v_Archery_g01_c03.avi", pts_unit="sec", output_format="TCHW") + + trans = transforms.Compose( + [ + ToTensorVideo(), + RandomHorizontalFlipVideo(), + UCFCenterCropVideo(512), + # NormalizeVideo(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True), + transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True), + ] + ) + + target_video_len = 32 + frame_interval = 1 + total_frames = len(vframes) + print(total_frames) + + temporal_sample = TemporalRandomCrop(target_video_len * frame_interval) + + # Sampling video frames + start_frame_ind, end_frame_ind = temporal_sample(total_frames) + # print(start_frame_ind) + # print(end_frame_ind) + assert end_frame_ind - start_frame_ind >= target_video_len + frame_indice = np.linspace(start_frame_ind, end_frame_ind - 1, target_video_len, dtype=int) + print(frame_indice) + + select_vframes = vframes[frame_indice] + print(select_vframes.shape) + print(select_vframes.dtype) + + select_vframes_trans = trans(select_vframes) + print(select_vframes_trans.shape) + print(select_vframes_trans.dtype) + + select_vframes_trans_int = ((select_vframes_trans * 0.5 + 0.5) * 255).to(dtype=torch.uint8) + print(select_vframes_trans_int.dtype) + print(select_vframes_trans_int.permute(0, 2, 3, 1).shape) + + io.write_video("./test.avi", select_vframes_trans_int.permute(0, 2, 3, 1), fps=8) + + for i in range(target_video_len): + save_image( + select_vframes_trans[i], os.path.join("./test000", "%04d.png" % i), normalize=True, value_range=(-1, 1) + ) diff --git a/opensora/models/__init__.py b/opensora/models/__init__.py new file mode 100644 index 00000000..60253499 --- /dev/null +++ b/opensora/models/__init__.py @@ -0,0 +1,6 @@ +from .dit import * +from .latte import * +from .pixart import * +from .stdit import * +from .text_encoder import * +from .vae import * diff --git a/opensora/models/dit/__init__.py b/opensora/models/dit/__init__.py new file mode 100644 index 00000000..94548a36 --- /dev/null +++ b/opensora/models/dit/__init__.py @@ -0,0 +1 @@ +from .dit import DiT, DiT_XL_2, DiT_XL_2x2 diff --git a/opensora/models/dit/dit.py b/opensora/models/dit/dit.py new file mode 100644 index 00000000..a23dd7b5 --- /dev/null +++ b/opensora/models/dit/dit.py @@ -0,0 +1,284 @@ +# Modified from Meta DiT + +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. +# -------------------------------------------------------- +# References: +# DiT: https://github.com/facebookresearch/DiT/tree/main +# GLIDE: https://github.com/openai/glide-text2im +# MAE: https://github.com/facebookresearch/mae/blob/main/models_mae.py +# -------------------------------------------------------- + +import numpy as np +import torch +import torch.nn as nn +import torch.utils.checkpoint +from einops import rearrange +from timm.models.vision_transformer import Mlp + +from opensora.acceleration.checkpoint import auto_grad_checkpoint +from opensora.models.layers.blocks import ( + Attention, + CaptionEmbedder, + FinalLayer, + LabelEmbedder, + PatchEmbed3D, + TimestepEmbedder, + approx_gelu, + get_1d_sincos_pos_embed, + get_2d_sincos_pos_embed, + get_layernorm, + modulate, +) +from opensora.registry import MODELS +from opensora.utils.ckpt_utils import load_checkpoint + + +class DiTBlock(nn.Module): + """ + A DiT block with adaptive layer norm zero (adaLN-Zero) conditioning. + """ + + def __init__( + self, + hidden_size, + num_heads, + mlp_ratio=4.0, + enable_flashattn=False, + enable_layernorm_kernel=False, + ): + super().__init__() + self.hidden_size = hidden_size + self.num_heads = num_heads + self.enable_flashattn = enable_flashattn + mlp_hidden_dim = int(hidden_size * mlp_ratio) + + self.norm1 = get_layernorm(hidden_size, eps=1e-6, affine=False, use_kernel=enable_layernorm_kernel) + self.attn = Attention( + hidden_size, + num_heads=num_heads, + qkv_bias=True, + enable_flashattn=enable_flashattn, + ) + self.norm2 = get_layernorm(hidden_size, eps=1e-6, affine=False, use_kernel=enable_layernorm_kernel) + self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, act_layer=approx_gelu, drop=0) + self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 6 * hidden_size, bias=True)) + + def forward(self, x, c): + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(6, dim=1) + x = x + gate_msa.unsqueeze(1) * self.attn(modulate(self.norm1, x, shift_msa, scale_msa)) + x = x + gate_mlp.unsqueeze(1) * self.mlp(modulate(self.norm2, x, shift_mlp, scale_mlp)) + return x + + +@MODELS.register_module() +class DiT(nn.Module): + """ + Diffusion model with a Transformer backbone. + """ + + def __init__( + self, + input_size=(16, 32, 32), + in_channels=4, + patch_size=(1, 2, 2), + hidden_size=1152, + depth=28, + num_heads=16, + mlp_ratio=4.0, + class_dropout_prob=0.1, + learn_sigma=True, + condition="text", + no_temporal_pos_emb=False, + caption_channels=512, + model_max_length=77, + dtype=torch.float32, + enable_flashattn=False, + enable_layernorm_kernel=False, + ): + super().__init__() + self.learn_sigma = learn_sigma + self.in_channels = in_channels + self.out_channels = in_channels * 2 if learn_sigma else in_channels + self.hidden_size = hidden_size + self.patch_size = patch_size + self.input_size = input_size + num_patches = np.prod([input_size[i] // patch_size[i] for i in range(3)]) + self.num_patches = num_patches + self.num_temporal = input_size[0] // patch_size[0] + self.num_spatial = num_patches // self.num_temporal + self.num_heads = num_heads + self.dtype = dtype + self.use_text_encoder = not condition.startswith("label") + if enable_flashattn: + assert dtype in [ + torch.float16, + torch.bfloat16, + ], f"Flash attention only supports float16 and bfloat16, but got {self.dtype}" + self.no_temporal_pos_emb = no_temporal_pos_emb + self.mlp_ratio = mlp_ratio + self.depth = depth + + self.register_buffer("pos_embed_spatial", self.get_spatial_pos_embed()) + self.register_buffer("pos_embed_temporal", self.get_temporal_pos_embed()) + + self.x_embedder = PatchEmbed3D(patch_size, in_channels, embed_dim=hidden_size) + if not self.use_text_encoder: + num_classes = int(condition.split("_")[-1]) + self.y_embedder = LabelEmbedder(num_classes, hidden_size, class_dropout_prob) + else: + self.y_embedder = CaptionEmbedder( + in_channels=caption_channels, + hidden_size=hidden_size, + uncond_prob=class_dropout_prob, + act_layer=approx_gelu, + token_num=1, # pooled token + ) + self.t_embedder = TimestepEmbedder(hidden_size) + self.blocks = nn.ModuleList( + [ + DiTBlock( + hidden_size, + num_heads, + mlp_ratio=mlp_ratio, + enable_flashattn=enable_flashattn, + enable_layernorm_kernel=enable_layernorm_kernel, + ) + for _ in range(depth) + ] + ) + self.final_layer = FinalLayer(hidden_size, np.prod(self.patch_size), self.out_channels) + + self.initialize_weights() + self.enable_flashattn = enable_flashattn + self.enable_layernorm_kernel = enable_layernorm_kernel + + def get_spatial_pos_embed(self): + pos_embed = get_2d_sincos_pos_embed( + self.hidden_size, + self.input_size[1] // self.patch_size[1], + ) + pos_embed = torch.from_numpy(pos_embed).float().unsqueeze(0).requires_grad_(False) + return pos_embed + + def get_temporal_pos_embed(self): + pos_embed = get_1d_sincos_pos_embed( + self.hidden_size, + self.input_size[0] // self.patch_size[0], + ) + pos_embed = torch.from_numpy(pos_embed).float().unsqueeze(0).requires_grad_(False) + return pos_embed + + def unpatchify(self, x): + c = self.out_channels + t, h, w = [self.input_size[i] // self.patch_size[i] for i in range(3)] + pt, ph, pw = self.patch_size + + x = x.reshape(shape=(x.shape[0], t, h, w, pt, ph, pw, c)) + x = rearrange(x, "n t h w r p q c -> n c t r h p w q") + imgs = x.reshape(shape=(x.shape[0], c, t * pt, h * ph, w * pw)) + return imgs + + def forward(self, x, t, y): + """ + Forward pass of DiT. + x: (B, C, T, H, W) tensor of inputs + t: (B,) tensor of diffusion timesteps + y: list of text + """ + # origin inputs should be float32, cast to specified dtype + x = x.to(self.dtype) + + # embedding + x = self.x_embedder(x) # (B, N, D) + x = rearrange(x, "b (t s) d -> b t s d", t=self.num_temporal, s=self.num_spatial) + x = x + self.pos_embed_spatial + if not self.no_temporal_pos_emb: + x = rearrange(x, "b t s d -> b s t d") + x = x + self.pos_embed_temporal + x = rearrange(x, "b s t d -> b (t s) d") + else: + x = rearrange(x, "b t s d -> b (t s) d") + + t = self.t_embedder(t, dtype=x.dtype) # (N, D) + y = self.y_embedder(y, self.training) # (N, D) + if self.use_text_encoder: + y = y.squeeze(1).squeeze(1) + condition = t + y + + # blocks + for _, block in enumerate(self.blocks): + c = condition + x = auto_grad_checkpoint(block, x, c) # (B, N, D) + + # final process + x = self.final_layer(x, condition) # (B, N, num_patches * out_channels) + x = self.unpatchify(x) # (B, out_channels, T, H, W) + + # cast to float32 for better accuracy + x = x.to(torch.float32) + return x + + def initialize_weights(self): + # Initialize transformer layers: + def _basic_init(module): + if isinstance(module, nn.Linear): + if module.weight.requires_grad_: + torch.nn.init.xavier_uniform_(module.weight) + if module.bias is not None: + nn.init.constant_(module.bias, 0) + + self.apply(_basic_init) + + # Initialize patch_embed like nn.Linear (instead of nn.Conv2d): + w = self.x_embedder.proj.weight.data + nn.init.xavier_uniform_(w.view([w.shape[0], -1])) + nn.init.constant_(self.x_embedder.proj.bias, 0) + + # Initialize timestep embedding MLP: + nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02) + nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02) + + # Zero-out adaLN modulation layers in DiT blocks: + for block in self.blocks: + nn.init.constant_(block.adaLN_modulation[-1].weight, 0) + nn.init.constant_(block.adaLN_modulation[-1].bias, 0) + + # Zero-out output layers: + nn.init.constant_(self.final_layer.adaLN_modulation[-1].weight, 0) + nn.init.constant_(self.final_layer.adaLN_modulation[-1].bias, 0) + nn.init.constant_(self.final_layer.linear.weight, 0) + nn.init.constant_(self.final_layer.linear.bias, 0) + + # Zero-out text embedding layers: + if self.use_text_encoder: + nn.init.normal_(self.y_embedder.y_proj.fc1.weight, std=0.02) + nn.init.normal_(self.y_embedder.y_proj.fc2.weight, std=0.02) + + +@MODELS.register_module("DiT-XL/2") +def DiT_XL_2(from_pretrained=None, **kwargs): + model = DiT( + depth=28, + hidden_size=1152, + patch_size=(1, 2, 2), + num_heads=16, + **kwargs, + ) + if from_pretrained is not None: + load_checkpoint(model, from_pretrained) + return model + + +@MODELS.register_module("DiT-XL/2x2") +def DiT_XL_2x2(from_pretrained=None, **kwargs): + model = DiT( + depth=28, + hidden_size=1152, + patch_size=(2, 2, 2), + num_heads=16, + **kwargs, + ) + if from_pretrained is not None: + load_checkpoint(model, from_pretrained) + return model diff --git a/opensora/models/latte/__init__.py b/opensora/models/latte/__init__.py new file mode 100644 index 00000000..f9d918ad --- /dev/null +++ b/opensora/models/latte/__init__.py @@ -0,0 +1 @@ +from .latte import Latte, Latte_XL_2, Latte_XL_2x2 diff --git a/opensora/models/latte/latte.py b/opensora/models/latte/latte.py new file mode 100644 index 00000000..3f8f9685 --- /dev/null +++ b/opensora/models/latte/latte.py @@ -0,0 +1,112 @@ +# Copyright 2024 Vchitect/Latte +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License.# Modified from Latte +# +# +# This file is mofied from https://github.com/Vchitect/Latte/blob/main/models/latte.py +# +# With references to: +# Latte: https://github.com/Vchitect/Latte +# DiT: https://github.com/facebookresearch/DiT/tree/main + + +import torch +from einops import rearrange, repeat + +from opensora.acceleration.checkpoint import auto_grad_checkpoint +from opensora.models.dit import DiT +from opensora.registry import MODELS +from opensora.utils.ckpt_utils import load_checkpoint + + +@MODELS.register_module() +class Latte(DiT): + def forward(self, x, t, y): + """ + Forward pass of DiT. + x: (B, C, T, H, W) tensor of inputs + t: (B,) tensor of diffusion timesteps + y: list of text + """ + # origin inputs should be float32, cast to specified dtype + x = x.to(self.dtype) + + # embedding + x = self.x_embedder(x) # (B, N, D) + x = rearrange(x, "b (t s) d -> b t s d", t=self.num_temporal, s=self.num_spatial) + x = x + self.pos_embed_spatial + x = rearrange(x, "b t s d -> b (t s) d") + + t = self.t_embedder(t, dtype=x.dtype) # (N, D) + y = self.y_embedder(y, self.training) # (N, D) + if self.use_text_encoder: + y = y.squeeze(1).squeeze(1) + condition = t + y + condition_spatial = repeat(condition, "b d -> (b t) d", t=self.num_temporal) + condition_temporal = repeat(condition, "b d -> (b s) d", s=self.num_spatial) + + # blocks + for i, block in enumerate(self.blocks): + if i % 2 == 0: + # spatial + x = rearrange(x, "b (t s) d -> (b t) s d", t=self.num_temporal, s=self.num_spatial) + c = condition_spatial + else: + # temporal + x = rearrange(x, "b (t s) d -> (b s) t d", t=self.num_temporal, s=self.num_spatial) + c = condition_temporal + if i == 1: + x = x + self.pos_embed_temporal + + x = auto_grad_checkpoint(block, x, c) # (B, N, D) + + if i % 2 == 0: + x = rearrange(x, "(b t) s d -> b (t s) d", t=self.num_temporal, s=self.num_spatial) + else: + x = rearrange(x, "(b s) t d -> b (t s) d", t=self.num_temporal, s=self.num_spatial) + + # final process + x = self.final_layer(x, condition) # (B, N, num_patches * out_channels) + x = self.unpatchify(x) # (B, out_channels, T, H, W) + + # cast to float32 for better accuracy + x = x.to(torch.float32) + return x + + +@MODELS.register_module("Latte-XL/2") +def Latte_XL_2(from_pretrained=None, **kwargs): + model = Latte( + depth=28, + hidden_size=1152, + patch_size=(1, 2, 2), + num_heads=16, + **kwargs, + ) + if from_pretrained is not None: + load_checkpoint(model, from_pretrained) + return model + + +@MODELS.register_module("Latte-XL/2x2") +def Latte_XL_2x2(from_pretrained=None, **kwargs): + model = Latte( + depth=28, + hidden_size=1152, + patch_size=(2, 2, 2), + num_heads=16, + **kwargs, + ) + if from_pretrained is not None: + load_checkpoint(model, from_pretrained) + return model diff --git a/opensora/models/layers/__init__.py b/opensora/models/layers/__init__.py new file mode 100644 index 00000000..e69de29b diff --git a/opensora/models/layers/blocks.py b/opensora/models/layers/blocks.py new file mode 100644 index 00000000..0f8bd596 --- /dev/null +++ b/opensora/models/layers/blocks.py @@ -0,0 +1,589 @@ +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. +# -------------------------------------------------------- +# References: +# PixArt: https://github.com/PixArt-alpha/PixArt-alpha +# Latte: https://github.com/Vchitect/Latte +# DiT: https://github.com/facebookresearch/DiT/tree/main +# GLIDE: https://github.com/openai/glide-text2im +# MAE: https://github.com/facebookresearch/mae/blob/main/models_mae.py +# -------------------------------------------------------- + +import math + +import numpy as np +import torch +import torch.distributed as dist +import torch.nn as nn +import torch.nn.functional as F +import torch.utils.checkpoint +import xformers.ops +from einops import rearrange +from timm.models.vision_transformer import Mlp + +from opensora.acceleration.communications import all_to_all, split_forward_gather_backward +from opensora.acceleration.parallel_states import get_sequence_parallel_group + +approx_gelu = lambda: nn.GELU(approximate="tanh") + + +def get_layernorm(hidden_size: torch.Tensor, eps: float, affine: bool, use_kernel: bool): + if use_kernel: + try: + from apex.normalization import FusedLayerNorm + + return FusedLayerNorm(hidden_size, elementwise_affine=affine, eps=eps) + except ImportError: + raise RuntimeError("FusedLayerNorm not available. Please install apex.") + else: + return nn.LayerNorm(hidden_size, eps, elementwise_affine=affine) + + +def modulate(norm_func, x, shift, scale): + # Suppose x is (B, N, D), shift is (B, D), scale is (B, D) + dtype = x.dtype + x = norm_func(x.to(torch.float32)).to(dtype) + x = x * (scale.unsqueeze(1) + 1) + shift.unsqueeze(1) + x = x.to(dtype) + return x + + +def t2i_modulate(x, shift, scale): + return x * (1 + scale) + shift + + +# =============================================== +# General-purpose Layers +# =============================================== + + +class PatchEmbed3D(nn.Module): + """Video to Patch Embedding. + + Args: + patch_size (int): Patch token size. Default: (2,4,4). + in_chans (int): Number of input video channels. Default: 3. + embed_dim (int): Number of linear projection output channels. Default: 96. + norm_layer (nn.Module, optional): Normalization layer. Default: None + """ + + def __init__( + self, + patch_size=(2, 4, 4), + in_chans=3, + embed_dim=96, + norm_layer=None, + flatten=True, + ): + super().__init__() + self.patch_size = patch_size + self.flatten = flatten + + self.in_chans = in_chans + self.embed_dim = embed_dim + + self.proj = nn.Conv3d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) + if norm_layer is not None: + self.norm = norm_layer(embed_dim) + else: + self.norm = None + + def forward(self, x): + """Forward function.""" + # padding + _, _, D, H, W = x.size() + if W % self.patch_size[2] != 0: + x = F.pad(x, (0, self.patch_size[2] - W % self.patch_size[2])) + if H % self.patch_size[1] != 0: + x = F.pad(x, (0, 0, 0, self.patch_size[1] - H % self.patch_size[1])) + if D % self.patch_size[0] != 0: + x = F.pad(x, (0, 0, 0, 0, 0, self.patch_size[0] - D % self.patch_size[0])) + + x = self.proj(x) # (B C T H W) + if self.norm is not None: + D, Wh, Ww = x.size(2), x.size(3), x.size(4) + x = x.flatten(2).transpose(1, 2) + x = self.norm(x) + x = x.transpose(1, 2).view(-1, self.embed_dim, D, Wh, Ww) + if self.flatten: + x = x.flatten(2).transpose(1, 2) # BCTHW -> BNC + return x + + +class Attention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int = 8, + qkv_bias: bool = False, + qk_norm: bool = False, + attn_drop: float = 0.0, + proj_drop: float = 0.0, + norm_layer: nn.Module = nn.LayerNorm, + enable_flashattn: bool = False, + ) -> None: + super().__init__() + assert dim % num_heads == 0, "dim should be divisible by num_heads" + self.dim = dim + self.num_heads = num_heads + self.head_dim = dim // num_heads + self.scale = self.head_dim**-0.5 + self.enable_flashattn = enable_flashattn + + self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) + self.q_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity() + self.k_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity() + self.attn_drop = nn.Dropout(attn_drop) + self.proj = nn.Linear(dim, dim) + self.proj_drop = nn.Dropout(proj_drop) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + B, N, C = x.shape + qkv = self.qkv(x) + qkv_shape = (B, N, 3, self.num_heads, self.head_dim) + if self.enable_flashattn: + qkv_permute_shape = (2, 0, 1, 3, 4) + else: + qkv_permute_shape = (2, 0, 3, 1, 4) + qkv = qkv.view(qkv_shape).permute(qkv_permute_shape) + q, k, v = qkv.unbind(0) + q, k = self.q_norm(q), self.k_norm(k) + if self.enable_flashattn: + from flash_attn import flash_attn_func + + x = flash_attn_func( + q, + k, + v, + dropout_p=self.attn_drop.p if self.training else 0.0, + softmax_scale=self.scale, + ) + else: + dtype = q.dtype + q = q * self.scale + attn = q @ k.transpose(-2, -1) # translate attn to float32 + attn = attn.to(torch.float32) + attn = attn.softmax(dim=-1) + attn = attn.to(dtype) # cast back attn to original dtype + attn = self.attn_drop(attn) + x = attn @ v + + x_output_shape = (B, N, C) + if not self.enable_flashattn: + x = x.transpose(1, 2) + x = x.reshape(x_output_shape) + x = self.proj(x) + x = self.proj_drop(x) + return x + + +class SeqParallelAttention(Attention): + def __init__( + self, + dim: int, + num_heads: int = 8, + qkv_bias: bool = False, + qk_norm: bool = False, + attn_drop: float = 0.0, + proj_drop: float = 0.0, + norm_layer: nn.Module = nn.LayerNorm, + enable_flashattn: bool = False, + ) -> None: + super().__init__( + dim=dim, + num_heads=num_heads, + qkv_bias=qkv_bias, + qk_norm=qk_norm, + attn_drop=attn_drop, + proj_drop=proj_drop, + norm_layer=norm_layer, + enable_flashattn=enable_flashattn, + ) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + B, N, C = x.shape # for sequence parallel here, the N is a local sequence length + qkv = self.qkv(x) + qkv_shape = (B, N, 3, self.num_heads, self.head_dim) + + qkv = qkv.view(qkv_shape) + + sp_group = get_sequence_parallel_group() + + # apply all_to_all to gather sequence and split attention heads + # [B, SUB_N, 3, NUM_HEAD, HEAD_DIM] -> [B, N, 3, NUM_HEAD_PER_DEVICE, HEAD_DIM] + qkv = all_to_all(qkv, sp_group, scatter_dim=3, gather_dim=1) + + if self.enable_flashattn: + qkv_permute_shape = (2, 0, 1, 3, 4) # [3, B, N, NUM_HEAD_PER_DEVICE, HEAD_DIM] + else: + qkv_permute_shape = (2, 0, 3, 1, 4) # [3, B, NUM_HEAD_PER_DEVICE, N, HEAD_DIM] + qkv = qkv.permute(qkv_permute_shape) + + q, k, v = qkv.unbind(0) + q, k = self.q_norm(q), self.k_norm(k) + if self.enable_flashattn: + from flash_attn import flash_attn_func + + x = flash_attn_func( + q, + k, + v, + dropout_p=self.attn_drop.p if self.training else 0.0, + softmax_scale=self.scale, + ) + else: + dtype = q.dtype + q = q * self.scale + attn = q @ k.transpose(-2, -1) # translate attn to float32 + attn = attn.to(torch.float32) + attn = attn.softmax(dim=-1) + attn = attn.to(dtype) # cast back attn to original dtype + attn = self.attn_drop(attn) + x = attn @ v + + if not self.enable_flashattn: + x = x.transpose(1, 2) + + # apply all to all to gather back attention heads and split sequence + # [B, N, NUM_HEAD_PER_DEVICE, HEAD_DIM] -> [B, SUB_N, NUM_HEAD, HEAD_DIM] + x = all_to_all(x, sp_group, scatter_dim=1, gather_dim=2) + + # reshape outputs back to [B, N, C] + x_output_shape = (B, N, C) + x = x.reshape(x_output_shape) + x = self.proj(x) + x = self.proj_drop(x) + return x + + +class MultiHeadCrossAttention(nn.Module): + def __init__(self, d_model, num_heads, attn_drop=0.0, proj_drop=0.0): + super(MultiHeadCrossAttention, self).__init__() + assert d_model % num_heads == 0, "d_model must be divisible by num_heads" + + self.d_model = d_model + self.num_heads = num_heads + self.head_dim = d_model // num_heads + + self.q_linear = nn.Linear(d_model, d_model) + self.kv_linear = nn.Linear(d_model, d_model * 2) + self.attn_drop = nn.Dropout(attn_drop) + self.proj = nn.Linear(d_model, d_model) + self.proj_drop = nn.Dropout(proj_drop) + + def forward(self, x, cond, mask=None): + # query/value: img tokens; key: condition; mask: if padding tokens + B, N, C = x.shape + + q = self.q_linear(x).view(1, -1, self.num_heads, self.head_dim) + kv = self.kv_linear(cond).view(1, -1, 2, self.num_heads, self.head_dim) + k, v = kv.unbind(2) + + attn_bias = None + if mask is not None: + attn_bias = xformers.ops.fmha.BlockDiagonalMask.from_seqlens([N] * B, mask) + x = xformers.ops.memory_efficient_attention(q, k, v, p=self.attn_drop.p, attn_bias=attn_bias) + + x = x.view(B, -1, C) + x = self.proj(x) + x = self.proj_drop(x) + return x + + +class SeqParallelMultiHeadCrossAttention(MultiHeadCrossAttention): + def __init__( + self, + d_model, + num_heads, + attn_drop=0.0, + proj_drop=0.0, + ): + super().__init__(d_model=d_model, num_heads=num_heads, attn_drop=attn_drop, proj_drop=proj_drop) + + def forward(self, x, cond, mask=None): + # query/value: img tokens; key: condition; mask: if padding tokens + sp_group = get_sequence_parallel_group() + sp_size = dist.get_world_size(sp_group) + B, SUB_N, C = x.shape + N = SUB_N * sp_size + + # shape: + # q, k, v: [B, SUB_N, NUM_HEADS, HEAD_DIM] + q = self.q_linear(x).view(B, -1, self.num_heads, self.head_dim) + kv = self.kv_linear(cond).view(B, -1, 2, self.num_heads, self.head_dim) + k, v = kv.unbind(2) + + # apply all_to_all to gather sequence and split attention heads + q = all_to_all(q, sp_group, scatter_dim=2, gather_dim=1) + + k = split_forward_gather_backward(k, get_sequence_parallel_group(), dim=2, grad_scale="down") + v = split_forward_gather_backward(v, get_sequence_parallel_group(), dim=2, grad_scale="down") + + q = q.view(1, -1, self.num_heads // sp_size, self.head_dim) + k = k.view(1, -1, self.num_heads // sp_size, self.head_dim) + v = v.view(1, -1, self.num_heads // sp_size, self.head_dim) + + # compute attention + attn_bias = None + if mask is not None: + attn_bias = xformers.ops.fmha.BlockDiagonalMask.from_seqlens([N] * B, mask) + x = xformers.ops.memory_efficient_attention(q, k, v, p=self.attn_drop.p, attn_bias=attn_bias) + + # apply all to all to gather back attention heads and scatter sequence + x = x.view(B, -1, self.num_heads // sp_size, self.head_dim) + x = all_to_all(x, sp_group, scatter_dim=1, gather_dim=2) + + # apply output projection + x = x.view(B, -1, C) + x = self.proj(x) + x = self.proj_drop(x) + return x + + +class FinalLayer(nn.Module): + """ + The final layer of DiT. + """ + + def __init__(self, hidden_size, num_patch, out_channels): + super().__init__() + self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) + self.linear = nn.Linear(hidden_size, num_patch * out_channels, bias=True) + self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True)) + + def forward(self, x, c): + shift, scale = self.adaLN_modulation(c).chunk(2, dim=1) + x = modulate(self.norm_final, x, shift, scale) + x = self.linear(x) + return x + + +class T2IFinalLayer(nn.Module): + """ + The final layer of PixArt. + """ + + def __init__(self, hidden_size, num_patch, out_channels): + super().__init__() + self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) + self.linear = nn.Linear(hidden_size, num_patch * out_channels, bias=True) + self.scale_shift_table = nn.Parameter(torch.randn(2, hidden_size) / hidden_size**0.5) + self.out_channels = out_channels + + def forward(self, x, t): + shift, scale = (self.scale_shift_table[None] + t[:, None]).chunk(2, dim=1) + x = t2i_modulate(self.norm_final(x), shift, scale) + x = self.linear(x) + return x + + +# =============================================== +# Embedding Layers for Timesteps and Class Labels +# =============================================== + + +class TimestepEmbedder(nn.Module): + """ + Embeds scalar timesteps into vector representations. + """ + + def __init__(self, hidden_size, frequency_embedding_size=256): + super().__init__() + self.mlp = nn.Sequential( + nn.Linear(frequency_embedding_size, hidden_size, bias=True), + nn.SiLU(), + nn.Linear(hidden_size, hidden_size, bias=True), + ) + self.frequency_embedding_size = frequency_embedding_size + + @staticmethod + def timestep_embedding(t, dim, max_period=10000): + """ + Create sinusoidal timestep embeddings. + :param t: a 1-D Tensor of N indices, one per batch element. + These may be fractional. + :param dim: the dimension of the output. + :param max_period: controls the minimum frequency of the embeddings. + :return: an (N, D) Tensor of positional embeddings. + """ + # https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py + half = dim // 2 + freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half) + freqs = freqs.to(device=t.device) + args = t[:, None].float() * freqs[None] + embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1) + if dim % 2: + embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1) + return embedding + + def forward(self, t, dtype): + t_freq = self.timestep_embedding(t, self.frequency_embedding_size) + if t_freq.dtype != dtype: + t_freq = t_freq.to(dtype) + t_emb = self.mlp(t_freq) + return t_emb + + +class LabelEmbedder(nn.Module): + """ + Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance. + """ + + def __init__(self, num_classes, hidden_size, dropout_prob): + super().__init__() + use_cfg_embedding = dropout_prob > 0 + self.embedding_table = nn.Embedding(num_classes + use_cfg_embedding, hidden_size) + self.num_classes = num_classes + self.dropout_prob = dropout_prob + + def token_drop(self, labels, force_drop_ids=None): + """ + Drops labels to enable classifier-free guidance. + """ + if force_drop_ids is None: + drop_ids = torch.rand(labels.shape[0]).cuda() < self.dropout_prob + else: + drop_ids = force_drop_ids == 1 + labels = torch.where(drop_ids, self.num_classes, labels) + return labels + + def forward(self, labels, train, force_drop_ids=None): + use_dropout = self.dropout_prob > 0 + if (train and use_dropout) or (force_drop_ids is not None): + labels = self.token_drop(labels, force_drop_ids) + return self.embedding_table(labels) + + +class SizeEmbedder(TimestepEmbedder): + """ + Embeds scalar timesteps into vector representations. + """ + + def __init__(self, hidden_size, frequency_embedding_size=256): + super().__init__(hidden_size=hidden_size, frequency_embedding_size=frequency_embedding_size) + self.mlp = nn.Sequential( + nn.Linear(frequency_embedding_size, hidden_size, bias=True), + nn.SiLU(), + nn.Linear(hidden_size, hidden_size, bias=True), + ) + self.frequency_embedding_size = frequency_embedding_size + self.outdim = hidden_size + + def forward(self, s, bs): + if s.ndim == 1: + s = s[:, None] + assert s.ndim == 2 + if s.shape[0] != bs: + s = s.repeat(bs // s.shape[0], 1) + assert s.shape[0] == bs + b, dims = s.shape[0], s.shape[1] + s = rearrange(s, "b d -> (b d)") + s_freq = self.timestep_embedding(s, self.frequency_embedding_size).to(self.dtype) + s_emb = self.mlp(s_freq) + s_emb = rearrange(s_emb, "(b d) d2 -> b (d d2)", b=b, d=dims, d2=self.outdim) + return s_emb + + @property + def dtype(self): + return next(self.parameters()).dtype + + +class CaptionEmbedder(nn.Module): + """ + Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance. + """ + + def __init__(self, in_channels, hidden_size, uncond_prob, act_layer=nn.GELU(approximate="tanh"), token_num=120): + super().__init__() + self.y_proj = Mlp( + in_features=in_channels, hidden_features=hidden_size, out_features=hidden_size, act_layer=act_layer, drop=0 + ) + self.register_buffer("y_embedding", nn.Parameter(torch.randn(token_num, in_channels) / in_channels**0.5)) + self.uncond_prob = uncond_prob + + def token_drop(self, caption, force_drop_ids=None): + """ + Drops labels to enable classifier-free guidance. + """ + if force_drop_ids is None: + drop_ids = torch.rand(caption.shape[0]).cuda() < self.uncond_prob + else: + drop_ids = force_drop_ids == 1 + caption = torch.where(drop_ids[:, None, None, None], self.y_embedding, caption) + return caption + + def forward(self, caption, train, force_drop_ids=None): + if train: + assert caption.shape[2:] == self.y_embedding.shape + use_dropout = self.uncond_prob > 0 + if (train and use_dropout) or (force_drop_ids is not None): + caption = self.token_drop(caption, force_drop_ids) + caption = self.y_proj(caption) + return caption + + +# =============================================== +# Sine/Cosine Positional Embedding Functions +# =============================================== +# https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py + + +def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=0, scale=1.0, base_size=None): + """ + grid_size: int of the grid height and width + return: + pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token) + """ + if not isinstance(grid_size, tuple): + grid_size = (grid_size, grid_size) + + grid_h = np.arange(grid_size[0], dtype=np.float32) / scale + grid_w = np.arange(grid_size[1], dtype=np.float32) / scale + if base_size is not None: + grid_h *= base_size / grid_size[0] + grid_w *= base_size / grid_size[1] + grid = np.meshgrid(grid_w, grid_h) # here w goes first + grid = np.stack(grid, axis=0) + + grid = grid.reshape([2, 1, grid_size[1], grid_size[0]]) + pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid) + if cls_token and extra_tokens > 0: + pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0) + return pos_embed + + +def get_2d_sincos_pos_embed_from_grid(embed_dim, grid): + assert embed_dim % 2 == 0 + + # use half of dimensions to encode grid_h + emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2) + emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2) + + emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D) + return emb + + +def get_1d_sincos_pos_embed(embed_dim, length, scale=1.0): + pos = np.arange(0, length)[..., None] / scale + return get_1d_sincos_pos_embed_from_grid(embed_dim, pos) + + +def get_1d_sincos_pos_embed_from_grid(embed_dim, pos): + """ + embed_dim: output dimension for each position + pos: a list of positions to be encoded: size (M,) + out: (M, D) + """ + assert embed_dim % 2 == 0 + omega = np.arange(embed_dim // 2, dtype=np.float64) + omega /= embed_dim / 2.0 + omega = 1.0 / 10000**omega # (D/2,) + + pos = pos.reshape(-1) # (M,) + out = np.einsum("m,d->md", pos, omega) # (M, D/2), outer product + + emb_sin = np.sin(out) # (M, D/2) + emb_cos = np.cos(out) # (M, D/2) + + emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D) + return emb diff --git a/opensora/models/pixart/__init__.py b/opensora/models/pixart/__init__.py new file mode 100644 index 00000000..cf832021 --- /dev/null +++ b/opensora/models/pixart/__init__.py @@ -0,0 +1 @@ +from .pixart import PixArt, PixArt_XL_2 diff --git a/opensora/models/pixart/pixart.py b/opensora/models/pixart/pixart.py new file mode 100644 index 00000000..849470ae --- /dev/null +++ b/opensora/models/pixart/pixart.py @@ -0,0 +1,389 @@ +# Adapted from PixArt +# +# Copyright (C) 2023 PixArt-alpha/PixArt-alpha +# +# This program is free software: you can redistribute it and/or modify +# it under the terms of the GNU Affero General Public License as published +# by the Free Software Foundation, either version 3 of the License, or +# (at your option) any later version. +# +# This program is distributed in the hope that it will be useful, +# but WITHOUT ANY WARRANTY; without even the implied warranty of +# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the +# GNU Affero General Public License for more details. +# +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. +# -------------------------------------------------------- +# References: +# PixArt: https://github.com/PixArt-alpha/PixArt-alpha +# DiT: https://github.com/facebookresearch/DiT/tree/main +# -------------------------------------------------------- + +import numpy as np +import torch +import torch.nn as nn +from einops import rearrange +from timm.models.layers import DropPath +from timm.models.vision_transformer import Mlp + +# from .builder import MODELS +from opensora.acceleration.checkpoint import auto_grad_checkpoint +from opensora.models.layers.blocks import ( + Attention, + CaptionEmbedder, + MultiHeadCrossAttention, + PatchEmbed3D, + SeqParallelAttention, + SeqParallelMultiHeadCrossAttention, + SizeEmbedder, + T2IFinalLayer, + TimestepEmbedder, + approx_gelu, + get_1d_sincos_pos_embed, + get_2d_sincos_pos_embed, + get_layernorm, + t2i_modulate, +) +from opensora.registry import MODELS +from opensora.utils.ckpt_utils import load_checkpoint + + +class PixArtBlock(nn.Module): + """ + A PixArt block with adaptive layer norm (adaLN-single) conditioning. + """ + + def __init__( + self, + hidden_size, + num_heads, + mlp_ratio=4.0, + drop_path=0.0, + enable_flashattn=False, + enable_layernorm_kernel=False, + enable_sequence_parallelism=False, + ): + super().__init__() + self.hidden_size = hidden_size + self.enable_flashattn = enable_flashattn + self._enable_sequence_parallelism = enable_sequence_parallelism + + if enable_sequence_parallelism: + self.attn_cls = SeqParallelAttention + self.mha_cls = SeqParallelMultiHeadCrossAttention + else: + self.attn_cls = Attention + self.mha_cls = MultiHeadCrossAttention + + self.norm1 = get_layernorm(hidden_size, eps=1e-6, affine=False, use_kernel=enable_layernorm_kernel) + self.attn = self.attn_cls( + hidden_size, + num_heads=num_heads, + qkv_bias=True, + enable_flashattn=enable_flashattn, + ) + self.cross_attn = self.mha_cls(hidden_size, num_heads) + self.norm2 = get_layernorm(hidden_size, eps=1e-6, affine=False, use_kernel=enable_layernorm_kernel) + self.mlp = Mlp( + in_features=hidden_size, hidden_features=int(hidden_size * mlp_ratio), act_layer=approx_gelu, drop=0 + ) + self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity() + self.scale_shift_table = nn.Parameter(torch.randn(6, hidden_size) / hidden_size**0.5) + + def forward(self, x, y, t, mask=None): + B, N, C = x.shape + + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ( + self.scale_shift_table[None] + t.reshape(B, 6, -1) + ).chunk(6, dim=1) + x = x + self.drop_path(gate_msa * self.attn(t2i_modulate(self.norm1(x), shift_msa, scale_msa)).reshape(B, N, C)) + x = x + self.cross_attn(x, y, mask) + x = x + self.drop_path(gate_mlp * self.mlp(t2i_modulate(self.norm2(x), shift_mlp, scale_mlp))) + + return x + + +@MODELS.register_module() +class PixArt(nn.Module): + """ + Diffusion model with a Transformer backbone. + """ + + def __init__( + self, + input_size=(1, 32, 32), + in_channels=4, + patch_size=(1, 2, 2), + hidden_size=1152, + depth=28, + num_heads=16, + mlp_ratio=4.0, + class_dropout_prob=0.1, + pred_sigma=True, + drop_path: float = 0.0, + no_temporal_pos_emb=False, + caption_channels=4096, + model_max_length=120, + dtype=torch.float32, + freeze=None, + space_scale=1.0, + time_scale=1.0, + enable_flashattn=False, + enable_layernorm_kernel=False, + ): + super().__init__() + self.pred_sigma = pred_sigma + self.in_channels = in_channels + self.out_channels = in_channels * 2 if pred_sigma else in_channels + self.hidden_size = hidden_size + self.patch_size = patch_size + self.input_size = input_size + num_patches = np.prod([input_size[i] // patch_size[i] for i in range(3)]) + self.num_patches = num_patches + self.num_temporal = input_size[0] // patch_size[0] + self.num_spatial = num_patches // self.num_temporal + self.base_size = int(np.sqrt(self.num_spatial)) + self.num_heads = num_heads + self.dtype = dtype + self.no_temporal_pos_emb = no_temporal_pos_emb + self.depth = depth + self.mlp_ratio = mlp_ratio + self.enable_flashattn = enable_flashattn + self.enable_layernorm_kernel = enable_layernorm_kernel + self.space_scale = space_scale + self.time_scale = time_scale + + self.x_embedder = PatchEmbed3D(patch_size, in_channels, hidden_size) + self.t_embedder = TimestepEmbedder(hidden_size) + self.t_block = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 6 * hidden_size, bias=True)) + self.y_embedder = CaptionEmbedder( + in_channels=caption_channels, + hidden_size=hidden_size, + uncond_prob=class_dropout_prob, + act_layer=approx_gelu, + token_num=model_max_length, + ) + + self.register_buffer("pos_embed", self.get_spatial_pos_embed()) + self.register_buffer("pos_embed_temporal", self.get_temporal_pos_embed()) + + drop_path = [x.item() for x in torch.linspace(0, drop_path, depth)] # stochastic depth decay rule + self.blocks = nn.ModuleList( + [ + PixArtBlock( + hidden_size, + num_heads, + mlp_ratio=mlp_ratio, + drop_path=drop_path[i], + enable_flashattn=enable_flashattn, + enable_layernorm_kernel=enable_layernorm_kernel, + ) + for i in range(depth) + ] + ) + self.final_layer = T2IFinalLayer(hidden_size, np.prod(self.patch_size), self.out_channels) + + self.initialize_weights() + if freeze is not None: + assert freeze in ["text"] + if freeze == "text": + self.freeze_text() + + def forward(self, x, timestep, y, mask=None): + """ + Forward pass of PixArt. + x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images) + t: (N,) tensor of diffusion timesteps + y: (N, 1, 120, C) tensor of class labels + """ + x = x.to(self.dtype) + timestep = timestep.to(self.dtype) + y = y.to(self.dtype) + + # embedding + x = self.x_embedder(x) # (B, N, D) + x = rearrange(x, "b (t s) d -> b t s d", t=self.num_temporal, s=self.num_spatial) + x = x + self.pos_embed + if not self.no_temporal_pos_emb: + x = rearrange(x, "b t s d -> b s t d") + x = x + self.pos_embed_temporal + x = rearrange(x, "b s t d -> b (t s) d") + else: + x = rearrange(x, "b t s d -> b (t s) d") + + t = self.t_embedder(timestep, dtype=x.dtype) # (N, D) + t0 = self.t_block(t) + y = self.y_embedder(y, self.training) # (N, 1, L, D) + if mask is not None: + if mask.shape[0] != y.shape[0]: + mask = mask.repeat(y.shape[0] // mask.shape[0], 1) + mask = mask.squeeze(1).squeeze(1) + y = y.squeeze(1).masked_select(mask.unsqueeze(-1) != 0).view(1, -1, x.shape[-1]) + y_lens = mask.sum(dim=1).tolist() + else: + y_lens = [y.shape[2]] * y.shape[0] + y = y.squeeze(1).view(1, -1, x.shape[-1]) + + # blocks + for block in self.blocks: + x = auto_grad_checkpoint(block, x, y, t0, y_lens) + + # final process + x = self.final_layer(x, t) # (N, T, patch_size ** 2 * out_channels) + x = self.unpatchify(x) # (N, out_channels, H, W) + + # cast to float32 for better accuracy + x = x.to(torch.float32) + return x + + def unpatchify(self, x): + c = self.out_channels + t, h, w = [self.input_size[i] // self.patch_size[i] for i in range(3)] + pt, ph, pw = self.patch_size + + x = x.reshape(shape=(x.shape[0], t, h, w, pt, ph, pw, c)) + x = rearrange(x, "n t h w r p q c -> n c t r h p w q") + imgs = x.reshape(shape=(x.shape[0], c, t * pt, h * ph, w * pw)) + return imgs + + def get_spatial_pos_embed(self, grid_size=None): + if grid_size is None: + grid_size = self.input_size[1:] + pos_embed = get_2d_sincos_pos_embed( + self.hidden_size, + (grid_size[0] // self.patch_size[1], grid_size[1] // self.patch_size[2]), + scale=self.space_scale, + base_size=self.base_size, + ) + pos_embed = torch.from_numpy(pos_embed).float().unsqueeze(0).requires_grad_(False) + return pos_embed + + def get_temporal_pos_embed(self): + pos_embed = get_1d_sincos_pos_embed( + self.hidden_size, + self.input_size[0] // self.patch_size[0], + scale=self.time_scale, + ) + pos_embed = torch.from_numpy(pos_embed).float().unsqueeze(0).requires_grad_(False) + return pos_embed + + def freeze_text(self): + for n, p in self.named_parameters(): + if "cross_attn" in n: + p.requires_grad = False + + def initialize_weights(self): + # Initialize transformer layers: + def _basic_init(module): + if isinstance(module, nn.Linear): + torch.nn.init.xavier_uniform_(module.weight) + if module.bias is not None: + nn.init.constant_(module.bias, 0) + + self.apply(_basic_init) + + # Initialize patch_embed like nn.Linear (instead of nn.Conv2d): + w = self.x_embedder.proj.weight.data + nn.init.xavier_uniform_(w.view([w.shape[0], -1])) + + # Initialize timestep embedding MLP: + nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02) + nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02) + nn.init.normal_(self.t_block[1].weight, std=0.02) + + # Initialize caption embedding MLP: + nn.init.normal_(self.y_embedder.y_proj.fc1.weight, std=0.02) + nn.init.normal_(self.y_embedder.y_proj.fc2.weight, std=0.02) + + # Zero-out adaLN modulation layers in PixArt blocks: + for block in self.blocks: + nn.init.constant_(block.cross_attn.proj.weight, 0) + nn.init.constant_(block.cross_attn.proj.bias, 0) + + # Zero-out output layers: + nn.init.constant_(self.final_layer.linear.weight, 0) + nn.init.constant_(self.final_layer.linear.bias, 0) + + +@MODELS.register_module() +class PixArtMS(PixArt): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + assert self.hidden_size % 3 == 0, "hidden_size must be divisible by 3" + self.csize_embedder = SizeEmbedder(self.hidden_size // 3) + self.ar_embedder = SizeEmbedder(self.hidden_size // 3) + + def forward(self, x, timestep, y, mask=None, data_info=None): + """ + Forward pass of PixArt. + x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images) + t: (N,) tensor of diffusion timesteps + y: (N, 1, 120, C) tensor of class labels + """ + x = x.to(self.dtype) + timestep = timestep.to(self.dtype) + y = y.to(self.dtype) + + c_size = data_info["hw"] + ar = data_info["ar"] + pos_embed = self.get_spatial_pos_embed((x.shape[-2], x.shape[-1])).to(x.dtype) + + # embedding + x = self.x_embedder(x) # (B, N, D) + x = rearrange(x, "b (t s) d -> b t s d", t=self.num_temporal, s=self.num_spatial) + x = x + pos_embed.to(x.device) + if not self.no_temporal_pos_emb: + x = rearrange(x, "b t s d -> b s t d") + x = x + self.pos_embed_temporal + x = rearrange(x, "b s t d -> b (t s) d") + else: + x = rearrange(x, "b t s d -> b (t s) d") + + t = self.t_embedder(timestep, dtype=x.dtype) # (N, D) + B = x.shape[0] + csize = self.csize_embedder(c_size, B) + ar = self.ar_embedder(ar, B) + t = t + torch.cat([csize, ar], dim=1) + + t0 = self.t_block(t) + y = self.y_embedder(y, self.training) # (N, 1, L, D) + if mask is not None: + if mask.shape[0] != y.shape[0]: + mask = mask.repeat(y.shape[0] // mask.shape[0], 1) + mask = mask.squeeze(1).squeeze(1) + y = y.squeeze(1).masked_select(mask.unsqueeze(-1) != 0).view(1, -1, x.shape[-1]) + y_lens = mask.sum(dim=1).tolist() + else: + y_lens = [y.shape[2]] * y.shape[0] + y = y.squeeze(1).view(1, -1, x.shape[-1]) + + # blocks + for block in self.blocks: + x = block(x, y, t0, y_lens) + + # final process + x = self.final_layer(x, t) # (N, T, patch_size ** 2 * out_channels) + x = self.unpatchify(x) # (N, out_channels, H, W) + + # cast to float32 for better accuracy + x = x.to(torch.float32) + return x + + +@MODELS.register_module("PixArt-XL/2") +def PixArt_XL_2(from_pretrained=None, **kwargs): + model = PixArt(depth=28, hidden_size=1152, patch_size=(1, 2, 2), num_heads=16, **kwargs) + if from_pretrained is not None: + load_checkpoint(model, from_pretrained) + return model + + +@MODELS.register_module("PixArtMS-XL/2") +def PixArtMS_XL_2(from_pretrained=None, **kwargs): + model = PixArtMS(depth=28, hidden_size=1152, patch_size=(1, 2, 2), num_heads=16, **kwargs) + if from_pretrained is not None: + load_checkpoint(model, from_pretrained) + return model diff --git a/opensora/models/stdit/__init__.py b/opensora/models/stdit/__init__.py new file mode 100644 index 00000000..5ca2cc91 --- /dev/null +++ b/opensora/models/stdit/__init__.py @@ -0,0 +1 @@ +from .stdit import STDiT diff --git a/opensora/models/stdit/stdit.py b/opensora/models/stdit/stdit.py new file mode 100644 index 00000000..68db6815 --- /dev/null +++ b/opensora/models/stdit/stdit.py @@ -0,0 +1,388 @@ +import numpy as np +import torch +import torch.distributed as dist +import torch.nn as nn +from einops import rearrange +from timm.models.layers import DropPath +from timm.models.vision_transformer import Mlp + +from opensora.acceleration.checkpoint import auto_grad_checkpoint +from opensora.acceleration.communications import gather_forward_split_backward, split_forward_gather_backward +from opensora.acceleration.parallel_states import get_sequence_parallel_group +from opensora.models.layers.blocks import ( + Attention, + CaptionEmbedder, + MultiHeadCrossAttention, + PatchEmbed3D, + SeqParallelAttention, + SeqParallelMultiHeadCrossAttention, + T2IFinalLayer, + TimestepEmbedder, + approx_gelu, + get_1d_sincos_pos_embed, + get_2d_sincos_pos_embed, + get_layernorm, + t2i_modulate, +) +from opensora.registry import MODELS +from opensora.utils.ckpt_utils import load_checkpoint + + +class STDiTBlock(nn.Module): + def __init__( + self, + hidden_size, + num_heads, + d_s=None, + d_t=None, + mlp_ratio=4.0, + drop_path=0.0, + enable_flashattn=False, + enable_layernorm_kernel=False, + enable_sequence_parallelism=False, + ): + super().__init__() + self.hidden_size = hidden_size + self.enable_flashattn = enable_flashattn + self._enable_sequence_parallelism = enable_sequence_parallelism + + if enable_sequence_parallelism: + self.attn_cls = SeqParallelAttention + self.mha_cls = SeqParallelMultiHeadCrossAttention + else: + self.attn_cls = Attention + self.mha_cls = MultiHeadCrossAttention + + self.norm1 = get_layernorm(hidden_size, eps=1e-6, affine=False, use_kernel=enable_layernorm_kernel) + self.attn = self.attn_cls( + hidden_size, + num_heads=num_heads, + qkv_bias=True, + enable_flashattn=enable_flashattn, + ) + self.cross_attn = self.mha_cls(hidden_size, num_heads) + self.norm2 = get_layernorm(hidden_size, eps=1e-6, affine=False, use_kernel=enable_layernorm_kernel) + self.mlp = Mlp( + in_features=hidden_size, hidden_features=int(hidden_size * mlp_ratio), act_layer=approx_gelu, drop=0 + ) + self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity() + self.scale_shift_table = nn.Parameter(torch.randn(6, hidden_size) / hidden_size**0.5) + + # temporal attention + self.d_s = d_s + self.d_t = d_t + + if self._enable_sequence_parallelism: + sp_size = dist.get_world_size(get_sequence_parallel_group()) + # make sure d_t is divisible by sp_size + assert d_t % sp_size == 0 + self.d_t = d_t // sp_size + + self.attn_temp = self.attn_cls( + hidden_size, + num_heads=num_heads, + qkv_bias=True, + enable_flashattn=self.enable_flashattn, + ) + + def forward(self, x, y, t, mask=None, tpe=None): + B, N, C = x.shape + + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ( + self.scale_shift_table[None] + t.reshape(B, 6, -1) + ).chunk(6, dim=1) + x_m = t2i_modulate(self.norm1(x), shift_msa, scale_msa) + + # spatial branch + x_s = rearrange(x_m, "B (T S) C -> (B T) S C", T=self.d_t, S=self.d_s) + x_s = self.attn(x_s) + x_s = rearrange(x_s, "(B T) S C -> B (T S) C", T=self.d_t, S=self.d_s) + x = x + self.drop_path(gate_msa * x_s) + + # temporal branch + x_t = rearrange(x, "B (T S) C -> (B S) T C", T=self.d_t, S=self.d_s) + if tpe is not None: + x_t = x_t + tpe + x_t = self.attn_temp(x_t) + x_t = rearrange(x_t, "(B S) T C -> B (T S) C", T=self.d_t, S=self.d_s) + x = x + self.drop_path(gate_msa * x_t) + + # cross attn + x = x + self.cross_attn(x, y, mask) + + # mlp + x = x + self.drop_path(gate_mlp * self.mlp(t2i_modulate(self.norm2(x), shift_mlp, scale_mlp))) + + return x + + +@MODELS.register_module() +class STDiT(nn.Module): + def __init__( + self, + input_size=(1, 32, 32), + in_channels=4, + patch_size=(1, 2, 2), + hidden_size=1152, + depth=28, + num_heads=16, + mlp_ratio=4.0, + class_dropout_prob=0.1, + pred_sigma=True, + drop_path=0.0, + no_temporal_pos_emb=False, + caption_channels=4096, + model_max_length=120, + dtype=torch.float32, + space_scale=1.0, + time_scale=1.0, + freeze=None, + enable_flashattn=False, + enable_layernorm_kernel=False, + enable_sequence_parallelism=False, + ): + super().__init__() + self.pred_sigma = pred_sigma + self.in_channels = in_channels + self.out_channels = in_channels * 2 if pred_sigma else in_channels + self.hidden_size = hidden_size + self.patch_size = patch_size + self.input_size = input_size + num_patches = np.prod([input_size[i] // patch_size[i] for i in range(3)]) + self.num_patches = num_patches + self.num_temporal = input_size[0] // patch_size[0] + self.num_spatial = num_patches // self.num_temporal + self.num_heads = num_heads + self.dtype = dtype + self.no_temporal_pos_emb = no_temporal_pos_emb + self.depth = depth + self.mlp_ratio = mlp_ratio + self.enable_flashattn = enable_flashattn + self.enable_layernorm_kernel = enable_layernorm_kernel + self.space_scale = space_scale + self.time_scale = time_scale + + self.register_buffer("pos_embed", self.get_spatial_pos_embed()) + self.register_buffer("pos_embed_temporal", self.get_temporal_pos_embed()) + + self.x_embedder = PatchEmbed3D(patch_size, in_channels, hidden_size) + self.t_embedder = TimestepEmbedder(hidden_size) + self.t_block = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 6 * hidden_size, bias=True)) + self.y_embedder = CaptionEmbedder( + in_channels=caption_channels, + hidden_size=hidden_size, + uncond_prob=class_dropout_prob, + act_layer=approx_gelu, + token_num=model_max_length, + ) + + drop_path = [x.item() for x in torch.linspace(0, drop_path, depth)] + self.blocks = nn.ModuleList( + [ + STDiTBlock( + self.hidden_size, + self.num_heads, + mlp_ratio=self.mlp_ratio, + drop_path=drop_path[i], + enable_flashattn=self.enable_flashattn, + enable_layernorm_kernel=self.enable_layernorm_kernel, + enable_sequence_parallelism=enable_sequence_parallelism, + d_t=self.num_temporal, + d_s=self.num_spatial, + ) + for i in range(self.depth) + ] + ) + self.final_layer = T2IFinalLayer(hidden_size, np.prod(self.patch_size), self.out_channels) + + # init model + self.initialize_weights() + self.initialize_temporal() + if freeze is not None: + assert freeze in ["not_temporal", "text"] + if freeze == "not_temporal": + self.freeze_not_temporal() + elif freeze == "text": + self.freeze_text() + + # sequence parallel related configs + self.enable_sequence_parallelism = enable_sequence_parallelism + if enable_sequence_parallelism: + self.sp_rank = dist.get_rank(get_sequence_parallel_group()) + else: + self.sp_rank = None + + def forward(self, x, timestep, y, mask=None): + """ + Forward pass of STDiT. + Args: + x (torch.Tensor): latent representation of video; of shape [B, C, T, H, W] + timestep (torch.Tensor): diffusion time steps; of shape [B] + y (torch.Tensor): representation of prompts; of shape [B, 1, N_token, C] + mask (torch.Tensor): mask for selecting prompt tokens; of shape [B, N_token] + + Returns: + x (torch.Tensor): output latent representation; of shape [B, C, T, H, W] + """ + + x = x.to(self.dtype) + timestep = timestep.to(self.dtype) + y = y.to(self.dtype) + + # embedding + x = self.x_embedder(x) # [B, N, C] + x = rearrange(x, "B (T S) C -> B T S C", T=self.num_temporal, S=self.num_spatial) + x = x + self.pos_embed + x = rearrange(x, "B T S C -> B (T S) C") + + # shard over the sequence dim if sp is enabled + if self.enable_sequence_parallelism: + x = split_forward_gather_backward(x, get_sequence_parallel_group(), dim=1, grad_scale="down") + + t = self.t_embedder(timestep, dtype=x.dtype) # [B, C] + t0 = self.t_block(t) # [B, C] + y = self.y_embedder(y, self.training) # [B, 1, N_token, C] + + if mask is not None: + if mask.shape[0] != y.shape[0]: + mask = mask.repeat(y.shape[0] // mask.shape[0], 1) + mask = mask.squeeze(1).squeeze(1) + y = y.squeeze(1).masked_select(mask.unsqueeze(-1) != 0).view(1, -1, x.shape[-1]) + y_lens = mask.sum(dim=1).tolist() + else: + y_lens = [y.shape[2]] * y.shape[0] + y = y.squeeze(1).view(1, -1, x.shape[-1]) + + # blocks + for i, block in enumerate(self.blocks): + if i == 0: + if self.enable_sequence_parallelism: + tpe = torch.chunk( + self.pos_embed_temporal, dist.get_world_size(get_sequence_parallel_group()), dim=1 + )[self.sp_rank].contiguous() + else: + tpe = self.pos_embed_temporal + else: + tpe = None + x = auto_grad_checkpoint(block, x, y, t0, y_lens, tpe) + + if self.enable_sequence_parallelism: + x = gather_forward_split_backward(x, get_sequence_parallel_group(), dim=1, grad_scale="up") + # x.shape: [B, N, C] + + # final process + x = self.final_layer(x, t) # [B, N, C=T_p * H_p * W_p * C_out] + x = self.unpatchify(x) # [B, C_out, T, H, W] + + # cast to float32 for better accuracy + x = x.to(torch.float32) + return x + + def unpatchify(self, x): + """ + Args: + x (torch.Tensor): of shape [B, N, C] + + Return: + x (torch.Tensor): of shape [B, C_out, T, H, W] + """ + + N_t, N_h, N_w = [self.input_size[i] // self.patch_size[i] for i in range(3)] + T_p, H_p, W_p = self.patch_size + x = rearrange( + x, + "B (N_t N_h N_w) (T_p H_p W_p C_out) -> B C_out (N_t T_p) (N_h H_p) (N_w W_p)", + N_t=N_t, + N_h=N_h, + N_w=N_w, + T_p=T_p, + H_p=H_p, + W_p=W_p, + C_out=self.out_channels, + ) + return x + + def unpatchify_old(self, x): + c = self.out_channels + t, h, w = [self.input_size[i] // self.patch_size[i] for i in range(3)] + pt, ph, pw = self.patch_size + + x = x.reshape(shape=(x.shape[0], t, h, w, pt, ph, pw, c)) + x = rearrange(x, "n t h w r p q c -> n c t r h p w q") + imgs = x.reshape(shape=(x.shape[0], c, t * pt, h * ph, w * pw)) + return imgs + + def get_spatial_pos_embed(self, grid_size=None): + if grid_size is None: + grid_size = self.input_size[1:] + pos_embed = get_2d_sincos_pos_embed( + self.hidden_size, + (grid_size[0] // self.patch_size[1], grid_size[1] // self.patch_size[2]), + scale=self.space_scale, + ) + pos_embed = torch.from_numpy(pos_embed).float().unsqueeze(0).requires_grad_(False) + return pos_embed + + def get_temporal_pos_embed(self): + pos_embed = get_1d_sincos_pos_embed( + self.hidden_size, + self.input_size[0] // self.patch_size[0], + scale=self.time_scale, + ) + pos_embed = torch.from_numpy(pos_embed).float().unsqueeze(0).requires_grad_(False) + return pos_embed + + def freeze_not_temporal(self): + for n, p in self.named_parameters(): + if "attn_temp" not in n: + p.requires_grad = False + + def freeze_text(self): + for n, p in self.named_parameters(): + if "cross_attn" in n: + p.requires_grad = False + + def initialize_temporal(self): + for block in self.blocks: + nn.init.constant_(block.attn_temp.proj.weight, 0) + nn.init.constant_(block.attn_temp.proj.bias, 0) + + def initialize_weights(self): + # Initialize transformer layers: + def _basic_init(module): + if isinstance(module, nn.Linear): + torch.nn.init.xavier_uniform_(module.weight) + if module.bias is not None: + nn.init.constant_(module.bias, 0) + + self.apply(_basic_init) + + # Initialize patch_embed like nn.Linear (instead of nn.Conv2d): + w = self.x_embedder.proj.weight.data + nn.init.xavier_uniform_(w.view([w.shape[0], -1])) + + # Initialize timestep embedding MLP: + nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02) + nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02) + nn.init.normal_(self.t_block[1].weight, std=0.02) + + # Initialize caption embedding MLP: + nn.init.normal_(self.y_embedder.y_proj.fc1.weight, std=0.02) + nn.init.normal_(self.y_embedder.y_proj.fc2.weight, std=0.02) + + # Zero-out adaLN modulation layers in PixArt blocks: + for block in self.blocks: + nn.init.constant_(block.cross_attn.proj.weight, 0) + nn.init.constant_(block.cross_attn.proj.bias, 0) + + # Zero-out output layers: + nn.init.constant_(self.final_layer.linear.weight, 0) + nn.init.constant_(self.final_layer.linear.bias, 0) + + +@MODELS.register_module("STDiT-XL/2") +def STDiT_XL_2(from_pretrained=None, **kwargs): + model = STDiT(depth=28, hidden_size=1152, patch_size=(1, 2, 2), num_heads=16, **kwargs) + if from_pretrained is not None: + load_checkpoint(model, from_pretrained) + return model diff --git a/opensora/models/text_encoder/__init__.py b/opensora/models/text_encoder/__init__.py new file mode 100644 index 00000000..9fc6a999 --- /dev/null +++ b/opensora/models/text_encoder/__init__.py @@ -0,0 +1,3 @@ +from .classes import ClassEncoder +from .clip import ClipEncoder +from .t5 import T5Encoder diff --git a/opensora/models/text_encoder/classes.py b/opensora/models/text_encoder/classes.py new file mode 100644 index 00000000..f02c9f29 --- /dev/null +++ b/opensora/models/text_encoder/classes.py @@ -0,0 +1,20 @@ +import torch + +from opensora.registry import MODELS + + +@MODELS.register_module("classes") +class ClassEncoder: + def __init__(self, num_classes, model_max_length=None, device="cuda", dtype=torch.float): + self.num_classes = num_classes + self.y_embedder = None + + self.model_max_length = model_max_length + self.output_dim = None + self.device = device + + def encode(self, text): + return dict(y=torch.tensor([int(t) for t in text]).to(self.device)) + + def null(self, n): + return torch.tensor([self.num_classes] * n).to(self.device) diff --git a/opensora/models/text_encoder/clip.py b/opensora/models/text_encoder/clip.py new file mode 100644 index 00000000..c628d02b --- /dev/null +++ b/opensora/models/text_encoder/clip.py @@ -0,0 +1,114 @@ +# Copyright 2024 Vchitect/Latte +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License.# Modified from Latte +# +# This file is adapted from the Latte project. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. +# -------------------------------------------------------- +# References: +# Latte: https://github.com/Vchitect/Latte +# DiT: https://github.com/facebookresearch/DiT/tree/main +# -------------------------------------------------------- + + +import torch +import torch.nn as nn +import transformers +from transformers import CLIPTextModel, CLIPTokenizer + +from opensora.registry import MODELS + +transformers.logging.set_verbosity_error() + + +class AbstractEncoder(nn.Module): + def __init__(self): + super().__init__() + + def encode(self, *args, **kwargs): + raise NotImplementedError + + +class FrozenCLIPEmbedder(AbstractEncoder): + """Uses the CLIP transformer encoder for text (from Hugging Face)""" + + def __init__(self, path="openai/clip-vit-huge-patch14", device="cuda", max_length=77): + super().__init__() + self.tokenizer = CLIPTokenizer.from_pretrained(path) + self.transformer = CLIPTextModel.from_pretrained(path) + self.device = device + self.max_length = max_length + self._freeze() + + def _freeze(self): + self.transformer = self.transformer.eval() + for param in self.parameters(): + param.requires_grad = False + + def forward(self, text): + batch_encoding = self.tokenizer( + text, + truncation=True, + max_length=self.max_length, + return_length=True, + return_overflowing_tokens=False, + padding="max_length", + return_tensors="pt", + ) + tokens = batch_encoding["input_ids"].to(self.device) + outputs = self.transformer(input_ids=tokens) + + z = outputs.last_hidden_state + pooled_z = outputs.pooler_output + return z, pooled_z + + def encode(self, text): + return self(text) + + +@MODELS.register_module("clip") +class ClipEncoder: + """ + Embeds text prompt into vector representations. Also handles text dropout for classifier-free guidance. + """ + + def __init__( + self, + from_pretrained, + model_max_length=77, + device="cuda", + dtype=torch.float, + ): + super().__init__() + assert from_pretrained is not None, "Please specify the path to the T5 model" + + self.text_encoder = FrozenCLIPEmbedder(path=from_pretrained, max_length=model_max_length).to(device, dtype) + self.y_embedder = None + + self.model_max_length = model_max_length + self.output_dim = self.text_encoder.transformer.config.hidden_size + + def encode(self, text): + _, pooled_embeddings = self.text_encoder.encode(text) + y = pooled_embeddings.unsqueeze(1).unsqueeze(1) + return dict(y=y) + + def null(self, n): + null_y = self.y_embedder.y_embedding[None].repeat(n, 1, 1)[:, None] + return null_y + + def to(self, dtype): + self.text_encoder = self.text_encoder.to(dtype) + return self diff --git a/opensora/models/text_encoder/t5.py b/opensora/models/text_encoder/t5.py new file mode 100644 index 00000000..f93612a7 --- /dev/null +++ b/opensora/models/text_encoder/t5.py @@ -0,0 +1,358 @@ +# Adapted from PixArt +# +# Copyright (C) 2023 PixArt-alpha/PixArt-alpha +# +# This program is free software: you can redistribute it and/or modify +# it under the terms of the GNU Affero General Public License as published +# by the Free Software Foundation, either version 3 of the License, or +# (at your option) any later version. +# +# This program is distributed in the hope that it will be useful, +# but WITHOUT ANY WARRANTY; without even the implied warranty of +# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the +# GNU Affero General Public License for more details. +# +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. +# -------------------------------------------------------- +# References: +# PixArt: https://github.com/PixArt-alpha/PixArt-alpha +# T5: https://github.com/google-research/text-to-text-transfer-transformer +# -------------------------------------------------------- + + +import html +import os +import re +import urllib.parse as ul + +import ftfy +import torch +from bs4 import BeautifulSoup +from huggingface_hub import hf_hub_download +from transformers import AutoTokenizer, T5EncoderModel + +from opensora.registry import MODELS + + +class T5Embedder: + available_models = ["t5-v1_1-xxl"] + bad_punct_regex = re.compile( + r"[" + "#®•©™&@·º½¾¿¡§~" + "\)" + "\(" + "\]" + "\[" + "\}" + "\{" + "\|" + "\\" + "\/" + "\*" + r"]{1,}" + ) # noqa + + def __init__( + self, + device, + dir_or_name="t5-v1_1-xxl", + *, + local_cache=False, + cache_dir=None, + hf_token=None, + use_text_preprocessing=True, + t5_model_kwargs=None, + torch_dtype=None, + use_offload_folder=None, + model_max_length=120, + ): + self.device = torch.device(device) + self.torch_dtype = torch_dtype or torch.bfloat16 + if t5_model_kwargs is None: + t5_model_kwargs = {"low_cpu_mem_usage": True, "torch_dtype": self.torch_dtype} + if use_offload_folder is not None: + t5_model_kwargs["offload_folder"] = use_offload_folder + t5_model_kwargs["device_map"] = { + "shared": self.device, + "encoder.embed_tokens": self.device, + "encoder.block.0": self.device, + "encoder.block.1": self.device, + "encoder.block.2": self.device, + "encoder.block.3": self.device, + "encoder.block.4": self.device, + "encoder.block.5": self.device, + "encoder.block.6": self.device, + "encoder.block.7": self.device, + "encoder.block.8": self.device, + "encoder.block.9": self.device, + "encoder.block.10": self.device, + "encoder.block.11": self.device, + "encoder.block.12": "disk", + "encoder.block.13": "disk", + "encoder.block.14": "disk", + "encoder.block.15": "disk", + "encoder.block.16": "disk", + "encoder.block.17": "disk", + "encoder.block.18": "disk", + "encoder.block.19": "disk", + "encoder.block.20": "disk", + "encoder.block.21": "disk", + "encoder.block.22": "disk", + "encoder.block.23": "disk", + "encoder.final_layer_norm": "disk", + "encoder.dropout": "disk", + } + else: + t5_model_kwargs["device_map"] = {"shared": self.device, "encoder": self.device} + + self.use_text_preprocessing = use_text_preprocessing + self.hf_token = hf_token + self.cache_dir = cache_dir or os.path.expanduser("~/.cache/IF_") + self.dir_or_name = dir_or_name + tokenizer_path, path = dir_or_name, dir_or_name + if local_cache: + cache_dir = os.path.join(self.cache_dir, dir_or_name) + tokenizer_path, path = cache_dir, cache_dir + elif dir_or_name in self.available_models: + cache_dir = os.path.join(self.cache_dir, dir_or_name) + for filename in [ + "config.json", + "special_tokens_map.json", + "spiece.model", + "tokenizer_config.json", + "pytorch_model.bin.index.json", + "pytorch_model-00001-of-00002.bin", + "pytorch_model-00002-of-00002.bin", + ]: + hf_hub_download( + repo_id=f"DeepFloyd/{dir_or_name}", + filename=filename, + cache_dir=cache_dir, + force_filename=filename, + token=self.hf_token, + ) + tokenizer_path, path = cache_dir, cache_dir + else: + cache_dir = os.path.join(self.cache_dir, "t5-v1_1-xxl") + for filename in [ + "config.json", + "special_tokens_map.json", + "spiece.model", + "tokenizer_config.json", + ]: + hf_hub_download( + repo_id="DeepFloyd/t5-v1_1-xxl", + filename=filename, + cache_dir=cache_dir, + force_filename=filename, + token=self.hf_token, + ) + tokenizer_path = cache_dir + + print(tokenizer_path) + self.tokenizer = AutoTokenizer.from_pretrained(tokenizer_path) + self.model = T5EncoderModel.from_pretrained(path, **t5_model_kwargs).eval() + self.model_max_length = model_max_length + + def get_text_embeddings(self, texts): + texts = [self.text_preprocessing(text) for text in texts] + + text_tokens_and_mask = self.tokenizer( + texts, + max_length=self.model_max_length, + padding="max_length", + truncation=True, + return_attention_mask=True, + add_special_tokens=True, + return_tensors="pt", + ) + + text_tokens_and_mask["input_ids"] = text_tokens_and_mask["input_ids"] + text_tokens_and_mask["attention_mask"] = text_tokens_and_mask["attention_mask"] + + with torch.no_grad(): + text_encoder_embs = self.model( + input_ids=text_tokens_and_mask["input_ids"].to(self.device), + attention_mask=text_tokens_and_mask["attention_mask"].to(self.device), + )["last_hidden_state"].detach() + return text_encoder_embs, text_tokens_and_mask["attention_mask"].to(self.device) + + def text_preprocessing(self, text): + if self.use_text_preprocessing: + # The exact text cleaning as was in the training stage: + text = self.clean_caption(text) + text = self.clean_caption(text) + return text + else: + return text.lower().strip() + + @staticmethod + def basic_clean(text): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + def clean_caption(self, caption): + caption = str(caption) + caption = ul.unquote_plus(caption) + caption = caption.strip().lower() + caption = re.sub("", "person", caption) + # urls: + caption = re.sub( + r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + caption = re.sub( + r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + # html: + caption = BeautifulSoup(caption, features="html.parser").text + + # @ + caption = re.sub(r"@[\w\d]+\b", "", caption) + + # 31C0—31EF CJK Strokes + # 31F0—31FF Katakana Phonetic Extensions + # 3200—32FF Enclosed CJK Letters and Months + # 3300—33FF CJK Compatibility + # 3400—4DBF CJK Unified Ideographs Extension A + # 4DC0—4DFF Yijing Hexagram Symbols + # 4E00—9FFF CJK Unified Ideographs + caption = re.sub(r"[\u31c0-\u31ef]+", "", caption) + caption = re.sub(r"[\u31f0-\u31ff]+", "", caption) + caption = re.sub(r"[\u3200-\u32ff]+", "", caption) + caption = re.sub(r"[\u3300-\u33ff]+", "", caption) + caption = re.sub(r"[\u3400-\u4dbf]+", "", caption) + caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption) + caption = re.sub(r"[\u4e00-\u9fff]+", "", caption) + ####################################################### + + # все виды тире / all types of dash --> "-" + caption = re.sub( + r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa + "-", + caption, + ) + + # кавычки к одному стандарту + caption = re.sub(r"[`´«»“”¨]", '"', caption) + caption = re.sub(r"[‘’]", "'", caption) + + # " + caption = re.sub(r""?", "", caption) + # & + caption = re.sub(r"&", "", caption) + + # ip adresses: + caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption) + + # article ids: + caption = re.sub(r"\d:\d\d\s+$", "", caption) + + # \n + caption = re.sub(r"\\n", " ", caption) + + # "#123" + caption = re.sub(r"#\d{1,3}\b", "", caption) + # "#12345.." + caption = re.sub(r"#\d{5,}\b", "", caption) + # "123456.." + caption = re.sub(r"\b\d{6,}\b", "", caption) + # filenames: + caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption) + + # + caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT""" + caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT""" + + caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT + caption = re.sub(r"\s+\.\s+", r" ", caption) # " . " + + # this-is-my-cute-cat / this_is_my_cute_cat + regex2 = re.compile(r"(?:\-|\_)") + if len(re.findall(regex2, caption)) > 3: + caption = re.sub(regex2, " ", caption) + + caption = self.basic_clean(caption) + + caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640 + caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc + caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231 + + caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption) + caption = re.sub(r"(free\s)?download(\sfree)?", "", caption) + caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption) + caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption) + caption = re.sub(r"\bpage\s+\d+\b", "", caption) + + caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a... + + caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption) + + caption = re.sub(r"\b\s+\:\s+", r": ", caption) + caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption) + caption = re.sub(r"\s+", " ", caption) + + caption.strip() + + caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption) + caption = re.sub(r"^[\'\_,\-\:;]", r"", caption) + caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption) + caption = re.sub(r"^\.\S+$", "", caption) + + return caption.strip() + + +@MODELS.register_module("t5") +class T5Encoder: + def __init__( + self, + from_pretrained=None, + model_max_length=120, + device="cuda", + dtype=torch.float, + local_cache=True, + shardformer=False, + ): + assert from_pretrained is not None, "Please specify the path to the T5 model" + + self.t5 = T5Embedder( + device=device, + torch_dtype=dtype, + local_cache=local_cache, + cache_dir=from_pretrained, + model_max_length=model_max_length, + ) + self.t5.model.to(dtype=dtype) + self.y_embedder = None + + self.model_max_length = model_max_length + self.output_dim = self.t5.model.config.d_model + + if shardformer: + self.shardformer_t5() + + def shardformer_t5(self): + from colossalai.shardformer import ShardConfig, ShardFormer + + from opensora.acceleration.shardformer.policy.t5_encoder import T5EncoderPolicy + from opensora.utils.misc import requires_grad + + shard_config = ShardConfig( + tensor_parallel_process_group=None, + pipeline_stage_manager=None, + enable_tensor_parallelism=False, + enable_fused_normalization=False, + enable_flash_attention=False, + enable_jit_fused=True, + enable_sequence_parallelism=False, + enable_sequence_overlap=False, + ) + shard_former = ShardFormer(shard_config=shard_config) + optim_model, _ = shard_former.optimize(self.t5.model, policy=T5EncoderPolicy()) + self.t5.model = optim_model.half() + + # ensure the weights are frozen + requires_grad(self.t5.model, False) + + def encode(self, text): + caption_embs, emb_masks = self.t5.get_text_embeddings(text) + caption_embs = caption_embs[:, None] + return dict(y=caption_embs, mask=emb_masks) + + def null(self, n): + null_y = self.y_embedder.y_embedding[None].repeat(n, 1, 1)[:, None] + return null_y diff --git a/opensora/models/vae/__init__.py b/opensora/models/vae/__init__.py new file mode 100644 index 00000000..63510b08 --- /dev/null +++ b/opensora/models/vae/__init__.py @@ -0,0 +1 @@ +from .vae import VideoAutoencoderKL, VideoAutoencoderKLTemporalDecoder diff --git a/opensora/models/vae/vae.py b/opensora/models/vae/vae.py new file mode 100644 index 00000000..363bbfed --- /dev/null +++ b/opensora/models/vae/vae.py @@ -0,0 +1,82 @@ +import torch +import torch.nn as nn +from diffusers.models import AutoencoderKL, AutoencoderKLTemporalDecoder +from einops import rearrange + +from opensora.registry import MODELS + + +@MODELS.register_module() +class VideoAutoencoderKL(nn.Module): + def __init__(self, from_pretrained=None, micro_batch_size=None): + super().__init__() + self.module = AutoencoderKL.from_pretrained(from_pretrained) + self.out_channels = self.module.config.latent_channels + self.patch_size = (1, 8, 8) + self.micro_batch_size = micro_batch_size + + def encode(self, x): + # x: (B, C, T, H, W) + B = x.shape[0] + x = rearrange(x, "B C T H W -> (B T) C H W") + + if self.micro_batch_size is None: + x = self.module.encode(x).latent_dist.sample().mul_(0.18215) + else: + bs = self.micro_batch_size + x_out = [] + for i in range(0, x.shape[0], bs): + x_bs = x[i : i + bs] + x_bs = self.module.encode(x_bs).latent_dist.sample().mul_(0.18215) + x_out.append(x_bs) + x = torch.cat(x_out, dim=0) + x = rearrange(x, "(B T) C H W -> B C T H W", B=B) + return x + + def decode(self, x): + # x: (B, C, T, H, W) + B = x.shape[0] + x = rearrange(x, "B C T H W -> (B T) C H W") + if self.micro_batch_size is None: + x = self.module.decode(x / 0.18215).sample + else: + bs = self.micro_batch_size + x_out = [] + for i in range(0, x.shape[0], bs): + x_bs = x[i : i + bs] + x_bs = self.module.decode(x_bs / 0.18215).sample + x_out.append(x_bs) + x = torch.cat(x_out, dim=0) + x = rearrange(x, "(B T) C H W -> B C T H W", B=B) + return x + + def get_latent_size(self, input_size): + for i in range(3): + assert input_size[i] % self.patch_size[i] == 0, "Input size must be divisible by patch size" + input_size = [input_size[i] // self.patch_size[i] for i in range(3)] + return input_size + + +@MODELS.register_module() +class VideoAutoencoderKLTemporalDecoder(nn.Module): + def __init__(self, from_pretrained=None): + super().__init__() + self.module = AutoencoderKLTemporalDecoder.from_pretrained(from_pretrained) + self.out_channels = self.module.config.latent_channels + self.patch_size = (1, 8, 8) + + def encode(self, x): + raise NotImplementedError + + def decode(self, x): + B, _, T = x.shape[:3] + x = rearrange(x, "B C T H W -> (B T) C H W") + x = self.module.decode(x / 0.18215, num_frames=T).sample + x = rearrange(x, "(B T) C H W -> B C T H W", B=B) + return x + + def get_latent_size(self, input_size): + for i in range(3): + assert input_size[i] % self.patch_size[i] == 0, "Input size must be divisible by patch size" + input_size = [input_size[i] // self.patch_size[i] for i in range(3)] + return input_size diff --git a/opensora/registry.py b/opensora/registry.py new file mode 100644 index 00000000..7797d36b --- /dev/null +++ b/opensora/registry.py @@ -0,0 +1,39 @@ +from copy import deepcopy + +import torch.nn as nn +from mmengine.registry import Registry + + +def build_module(module, builder, **kwargs): + """Build module from config or return the module itself. + + Args: + module (Union[dict, nn.Module]): The module to build. + builder (Registry): The registry to build module. + *args, **kwargs: Arguments passed to build function. + + Returns: + Any: The built module. + """ + if isinstance(module, dict): + cfg = deepcopy(module) + for k, v in kwargs.items(): + cfg[k] = v + return builder.build(cfg) + elif isinstance(module, nn.Module): + return module + elif module is None: + return None + else: + raise TypeError(f"Only support dict and nn.Module, but got {type(module)}.") + + +MODELS = Registry( + "model", + locations=["opensora.models"], +) + +SCHEDULERS = Registry( + "scheduler", + locations=["opensora.schedulers"], +) diff --git a/opensora/schedulers/__init__.py b/opensora/schedulers/__init__.py new file mode 100644 index 00000000..97ea76f9 --- /dev/null +++ b/opensora/schedulers/__init__.py @@ -0,0 +1,2 @@ +from .dpms import DPMS +from .iddpm import IDDPM diff --git a/opensora/schedulers/dpms/__init__.py b/opensora/schedulers/dpms/__init__.py new file mode 100644 index 00000000..f0cebbcd --- /dev/null +++ b/opensora/schedulers/dpms/__init__.py @@ -0,0 +1,50 @@ +from functools import partial + +import torch + +from opensora.registry import SCHEDULERS + +from .dpm_solver import DPMS + + +@SCHEDULERS.register_module("dpm-solver") +class DMP_SOLVER: + def __init__(self, num_sampling_steps=None, cfg_scale=4.0): + self.num_sampling_steps = num_sampling_steps + self.cfg_scale = cfg_scale + + def sample( + self, + model, + text_encoder, + z_size, + prompts, + device, + additional_args=None, + ): + n = len(prompts) + z = torch.randn(n, *z_size, device=device) + model_args = text_encoder.encode(prompts) + y = model_args.pop("y") + null_y = text_encoder.null(n) + if additional_args is not None: + model_args.update(additional_args) + + dpms = DPMS( + partial(forward_with_dpmsolver, model), + condition=y, + uncondition=null_y, + cfg_scale=self.cfg_scale, + model_kwargs=model_args, + ) + samples = dpms.sample(z, steps=self.num_sampling_steps, order=2, skip_type="time_uniform", method="multistep") + return samples + + +def forward_with_dpmsolver(self, x, timestep, y, **kwargs): + """ + dpm solver donnot need variance prediction + """ + # https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb + model_out = self.forward(x, timestep, y, **kwargs) + return model_out.chunk(2, dim=1)[0] diff --git a/opensora/schedulers/dpms/dpm_solver.py b/opensora/schedulers/dpms/dpm_solver.py new file mode 100644 index 00000000..106e59ec --- /dev/null +++ b/opensora/schedulers/dpms/dpm_solver.py @@ -0,0 +1,1570 @@ +# MIT License +# +# Copyright (c) 2022 Cheng Lu +# +# Permission is hereby granted, free of charge, to any person obtaining a copy +# of this software and associated documentation files (the "Software"), to deal +# in the Software without restriction, including without limitation the rights +# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +# copies of the Software, and to permit persons to whom the Software is +# furnished to do so, subject to the following conditions: +# +# +# This file is adapted from the dpm-solver project +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. +# -------------------------------------------------------- +# References: +# PixArt: https://github.com/PixArt-alpha/PixArt-alpha +# dpm-solver: https://github.com/LuChengTHU/dpm-solver +# -------------------------------------------------------- + +import math + +import numpy as np +import torch +from tqdm import tqdm + + +def _warmup_beta(beta_start, beta_end, num_diffusion_timesteps, warmup_frac): + betas = beta_end * np.ones(num_diffusion_timesteps, dtype=np.float64) + warmup_time = int(num_diffusion_timesteps * warmup_frac) + betas[:warmup_time] = np.linspace(beta_start, beta_end, warmup_time, dtype=np.float64) + return betas + + +def get_beta_schedule(beta_schedule, *, beta_start, beta_end, num_diffusion_timesteps): + """ + This is the deprecated API for creating beta schedules. + See get_named_beta_schedule() for the new library of schedules. + """ + if beta_schedule == "quad": + betas = ( + np.linspace( + beta_start**0.5, + beta_end**0.5, + num_diffusion_timesteps, + dtype=np.float64, + ) + ** 2 + ) + elif beta_schedule == "linear": + betas = np.linspace(beta_start, beta_end, num_diffusion_timesteps, dtype=np.float64) + elif beta_schedule == "warmup10": + betas = _warmup_beta(beta_start, beta_end, num_diffusion_timesteps, 0.1) + elif beta_schedule == "warmup50": + betas = _warmup_beta(beta_start, beta_end, num_diffusion_timesteps, 0.5) + elif beta_schedule == "const": + betas = beta_end * np.ones(num_diffusion_timesteps, dtype=np.float64) + elif beta_schedule == "jsd": # 1/T, 1/(T-1), 1/(T-2), ..., 1 + betas = 1.0 / np.linspace(num_diffusion_timesteps, 1, num_diffusion_timesteps, dtype=np.float64) + else: + raise NotImplementedError(beta_schedule) + assert betas.shape == (num_diffusion_timesteps,) + return betas + + +def get_named_beta_schedule(schedule_name, num_diffusion_timesteps): + """ + Get a pre-defined beta schedule for the given name. + The beta schedule library consists of beta schedules which remain similar + in the limit of num_diffusion_timesteps. + Beta schedules may be added, but should not be removed or changed once + they are committed to maintain backwards compatibility. + """ + if schedule_name == "linear": + # Linear schedule from Ho et al, extended to work for any number of + # diffusion steps. + scale = 1000 / num_diffusion_timesteps + return get_beta_schedule( + "linear", + beta_start=scale * 0.0001, + beta_end=scale * 0.02, + num_diffusion_timesteps=num_diffusion_timesteps, + ) + elif schedule_name == "squaredcos_cap_v2": + return betas_for_alpha_bar( + num_diffusion_timesteps, + lambda t: math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2, + ) + else: + raise NotImplementedError(f"unknown beta schedule: {schedule_name}") + + +def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, + which defines the cumulative product of (1-beta) over time from t = [0,1]. + :param num_diffusion_timesteps: the number of betas to produce. + :param alpha_bar: a lambda that takes an argument t from 0 to 1 and + produces the cumulative product of (1-beta) up to that + part of the diffusion process. + :param max_beta: the maximum beta to use; use values lower than 1 to + prevent singularities. + """ + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta)) + return np.array(betas) + + +class NoiseScheduleVP: + def __init__( + self, + schedule="discrete", + betas=None, + alphas_cumprod=None, + continuous_beta_0=0.1, + continuous_beta_1=20.0, + dtype=torch.float32, + ): + """Create a wrapper class for the forward SDE (VP type). + + *** + Update: We support discrete-time diffusion models by implementing a picewise linear interpolation for log_alpha_t. + We recommend to use schedule='discrete' for the discrete-time diffusion models, especially for high-resolution images. + *** + + The forward SDE ensures that the condition distribution q_{t|0}(x_t | x_0) = N ( alpha_t * x_0, sigma_t^2 * I ). + We further define lambda_t = log(alpha_t) - log(sigma_t), which is the half-logSNR (described in the DPM-Solver paper). + Therefore, we implement the functions for computing alpha_t, sigma_t and lambda_t. For t in [0, T], we have: + + log_alpha_t = self.marginal_log_mean_coeff(t) + sigma_t = self.marginal_std(t) + lambda_t = self.marginal_lambda(t) + + Moreover, as lambda(t) is an invertible function, we also support its inverse function: + + t = self.inverse_lambda(lambda_t) + + =============================================================== + + We support both discrete-time DPMs (trained on n = 0, 1, ..., N-1) and continuous-time DPMs (trained on t in [t_0, T]). + + 1. For discrete-time DPMs: + + For discrete-time DPMs trained on n = 0, 1, ..., N-1, we convert the discrete steps to continuous time steps by: + t_i = (i + 1) / N + e.g. for N = 1000, we have t_0 = 1e-3 and T = t_{N-1} = 1. + We solve the corresponding diffusion ODE from time T = 1 to time t_0 = 1e-3. + + Args: + betas: A `torch.Tensor`. The beta array for the discrete-time DPM. (See the original DDPM paper for details) + alphas_cumprod: A `torch.Tensor`. The cumprod alphas for the discrete-time DPM. (See the original DDPM paper for details) + + Note that we always have alphas_cumprod = cumprod(1 - betas). Therefore, we only need to set one of `betas` and `alphas_cumprod`. + + **Important**: Please pay special attention for the args for `alphas_cumprod`: + The `alphas_cumprod` is the \hat{alpha_n} arrays in the notations of DDPM. Specifically, DDPMs assume that + q_{t_n | 0}(x_{t_n} | x_0) = N ( \sqrt{\hat{alpha_n}} * x_0, (1 - \hat{alpha_n}) * I ). + Therefore, the notation \hat{alpha_n} is different from the notation alpha_t in DPM-Solver. In fact, we have + alpha_{t_n} = \sqrt{\hat{alpha_n}}, + and + log(alpha_{t_n}) = 0.5 * log(\hat{alpha_n}). + + + 2. For continuous-time DPMs: + + We support the linear VPSDE for the continuous time setting. The hyperparameters for the noise + schedule are the default settings in Yang Song's ScoreSDE: + + Args: + beta_min: A `float` number. The smallest beta for the linear schedule. + beta_max: A `float` number. The largest beta for the linear schedule. + T: A `float` number. The ending time of the forward process. + + =============================================================== + + Args: + schedule: A `str`. The noise schedule of the forward SDE. 'discrete' for discrete-time DPMs, + 'linear' for continuous-time DPMs. + Returns: + A wrapper object of the forward SDE (VP type). + + =============================================================== + + Example: + + # For discrete-time DPMs, given betas (the beta array for n = 0, 1, ..., N - 1): + >>> ns = NoiseScheduleVP('discrete', betas=betas) + + # For discrete-time DPMs, given alphas_cumprod (the \hat{alpha_n} array for n = 0, 1, ..., N - 1): + >>> ns = NoiseScheduleVP('discrete', alphas_cumprod=alphas_cumprod) + + # For continuous-time DPMs (VPSDE), linear schedule: + >>> ns = NoiseScheduleVP('linear', continuous_beta_0=0.1, continuous_beta_1=20.) + + """ + + if schedule not in ["discrete", "linear"]: + raise ValueError(f"Unsupported noise schedule {schedule}. The schedule needs to be 'discrete' or 'linear'") + + self.schedule = schedule + if schedule == "discrete": + if betas is not None: + log_alphas = 0.5 * torch.log(1 - betas).cumsum(dim=0) + else: + assert alphas_cumprod is not None + log_alphas = 0.5 * torch.log(alphas_cumprod) + self.T = 1.0 + self.log_alpha_array = ( + self.numerical_clip_alpha(log_alphas) + .reshape( + ( + 1, + -1, + ) + ) + .to(dtype=dtype) + ) + self.total_N = self.log_alpha_array.shape[1] + self.t_array = torch.linspace(0.0, 1.0, self.total_N + 1)[1:].reshape((1, -1)).to(dtype=dtype) + else: + self.T = 1.0 + self.total_N = 1000 + self.beta_0 = continuous_beta_0 + self.beta_1 = continuous_beta_1 + + def numerical_clip_alpha(self, log_alphas, clipped_lambda=-5.1): + """ + For some beta schedules such as cosine schedule, the log-SNR has numerical isssues. + We clip the log-SNR near t=T within -5.1 to ensure the stability. + Such a trick is very useful for diffusion models with the cosine schedule, such as i-DDPM, guided-diffusion and GLIDE. + """ + log_sigmas = 0.5 * torch.log(1.0 - torch.exp(2.0 * log_alphas)) + lambs = log_alphas - log_sigmas + idx = torch.searchsorted(torch.flip(lambs, [0]), clipped_lambda) + if idx > 0: + log_alphas = log_alphas[:-idx] + return log_alphas + + def marginal_log_mean_coeff(self, t): + """ + Compute log(alpha_t) of a given continuous-time label t in [0, T]. + """ + if self.schedule == "discrete": + return interpolate_fn( + t.reshape((-1, 1)), self.t_array.to(t.device), self.log_alpha_array.to(t.device) + ).reshape((-1)) + elif self.schedule == "linear": + return -0.25 * t**2 * (self.beta_1 - self.beta_0) - 0.5 * t * self.beta_0 + + def marginal_alpha(self, t): + """ + Compute alpha_t of a given continuous-time label t in [0, T]. + """ + return torch.exp(self.marginal_log_mean_coeff(t)) + + def marginal_std(self, t): + """ + Compute sigma_t of a given continuous-time label t in [0, T]. + """ + return torch.sqrt(1.0 - torch.exp(2.0 * self.marginal_log_mean_coeff(t))) + + def marginal_lambda(self, t): + """ + Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T]. + """ + log_mean_coeff = self.marginal_log_mean_coeff(t) + log_std = 0.5 * torch.log(1.0 - torch.exp(2.0 * log_mean_coeff)) + return log_mean_coeff - log_std + + def inverse_lambda(self, lamb): + """ + Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t. + """ + if self.schedule == "linear": + tmp = 2.0 * (self.beta_1 - self.beta_0) * torch.logaddexp(-2.0 * lamb, torch.zeros((1,)).to(lamb)) + Delta = self.beta_0**2 + tmp + return tmp / (torch.sqrt(Delta) + self.beta_0) / (self.beta_1 - self.beta_0) + elif self.schedule == "discrete": + log_alpha = -0.5 * torch.logaddexp(torch.zeros((1,)).to(lamb.device), -2.0 * lamb) + t = interpolate_fn( + log_alpha.reshape((-1, 1)), + torch.flip(self.log_alpha_array.to(lamb.device), [1]), + torch.flip(self.t_array.to(lamb.device), [1]), + ) + return t.reshape((-1,)) + + +def model_wrapper( + model, + noise_schedule, + model_type="noise", + model_kwargs={}, + guidance_type="uncond", + condition=None, + unconditional_condition=None, + guidance_scale=1.0, + classifier_fn=None, + classifier_kwargs={}, +): + """Create a wrapper function for the noise prediction model. + + DPM-Solver needs to solve the continuous-time diffusion ODEs. For DPMs trained on discrete-time labels, we need to + firstly wrap the model function to a noise prediction model that accepts the continuous time as the input. + + We support four types of the diffusion model by setting `model_type`: + + 1. "noise": noise prediction model. (Trained by predicting noise). + + 2. "x_start": data prediction model. (Trained by predicting the data x_0 at time 0). + + 3. "v": velocity prediction model. (Trained by predicting the velocity). + The "v" prediction is derivation detailed in Appendix D of [1], and is used in Imagen-Video [2]. + + [1] Salimans, Tim, and Jonathan Ho. "Progressive distillation for fast sampling of diffusion models." + arXiv preprint arXiv:2202.00512 (2022). + [2] Ho, Jonathan, et al. "Imagen Video: High Definition Video Generation with Diffusion Models." + arXiv preprint arXiv:2210.02303 (2022). + + 4. "score": marginal score function. (Trained by denoising score matching). + Note that the score function and the noise prediction model follows a simple relationship: + ``` + noise(x_t, t) = -sigma_t * score(x_t, t) + ``` + + We support three types of guided sampling by DPMs by setting `guidance_type`: + 1. "uncond": unconditional sampling by DPMs. + The input `model` has the following format: + `` + model(x, t_input, **model_kwargs) -> noise | x_start | v | score + `` + + 2. "classifier": classifier guidance sampling [3] by DPMs and another classifier. + The input `model` has the following format: + `` + model(x, t_input, **model_kwargs) -> noise | x_start | v | score + `` + + The input `classifier_fn` has the following format: + `` + classifier_fn(x, t_input, cond, **classifier_kwargs) -> logits(x, t_input, cond) + `` + + [3] P. Dhariwal and A. Q. Nichol, "Diffusion models beat GANs on image synthesis," + in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 8780-8794. + + 3. "classifier-free": classifier-free guidance sampling by conditional DPMs. + The input `model` has the following format: + `` + model(x, t_input, cond, **model_kwargs) -> noise | x_start | v | score + `` + And if cond == `unconditional_condition`, the model output is the unconditional DPM output. + + [4] Ho, Jonathan, and Tim Salimans. "Classifier-free diffusion guidance." + arXiv preprint arXiv:2207.12598 (2022). + + + The `t_input` is the time label of the model, which may be discrete-time labels (i.e. 0 to 999) + or continuous-time labels (i.e. epsilon to T). + + We wrap the model function to accept only `x` and `t_continuous` as inputs, and outputs the predicted noise: + `` + def model_fn(x, t_continuous) -> noise: + t_input = get_model_input_time(t_continuous) + return noise_pred(model, x, t_input, **model_kwargs) + `` + where `t_continuous` is the continuous time labels (i.e. epsilon to T). And we use `model_fn` for DPM-Solver. + + =============================================================== + + Args: + model: A diffusion model with the corresponding format described above. + noise_schedule: A noise schedule object, such as NoiseScheduleVP. + model_type: A `str`. The parameterization type of the diffusion model. + "noise" or "x_start" or "v" or "score". + model_kwargs: A `dict`. A dict for the other inputs of the model function. + guidance_type: A `str`. The type of the guidance for sampling. + "uncond" or "classifier" or "classifier-free". + condition: A pytorch tensor. The condition for the guided sampling. + Only used for "classifier" or "classifier-free" guidance type. + unconditional_condition: A pytorch tensor. The condition for the unconditional sampling. + Only used for "classifier-free" guidance type. + guidance_scale: A `float`. The scale for the guided sampling. + classifier_fn: A classifier function. Only used for the classifier guidance. + classifier_kwargs: A `dict`. A dict for the other inputs of the classifier function. + Returns: + A noise prediction model that accepts the noised data and the continuous time as the inputs. + """ + + def get_model_input_time(t_continuous): + """ + Convert the continuous-time `t_continuous` (in [epsilon, T]) to the model input time. + For discrete-time DPMs, we convert `t_continuous` in [1 / N, 1] to `t_input` in [0, 1000 * (N - 1) / N]. + For continuous-time DPMs, we just use `t_continuous`. + """ + if noise_schedule.schedule == "discrete": + return (t_continuous - 1.0 / noise_schedule.total_N) * 1000.0 + else: + return t_continuous + + def noise_pred_fn(x, t_continuous, cond=None): + t_input = get_model_input_time(t_continuous) + if cond is None: + output = model(x, t_input, **model_kwargs) + else: + output = model(x, t_input, cond, **model_kwargs) + if model_type == "noise": + return output + elif model_type == "x_start": + alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous) + return (x - expand_dims(alpha_t, x.dim()) * output) / expand_dims(sigma_t, x.dim()) + elif model_type == "v": + alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous) + return expand_dims(alpha_t, x.dim()) * output + expand_dims(sigma_t, x.dim()) * x + elif model_type == "score": + sigma_t = noise_schedule.marginal_std(t_continuous) + return -expand_dims(sigma_t, x.dim()) * output + + def cond_grad_fn(x, t_input): + """ + Compute the gradient of the classifier, i.e. nabla_{x} log p_t(cond | x_t). + """ + with torch.enable_grad(): + x_in = x.detach().requires_grad_(True) + log_prob = classifier_fn(x_in, t_input, condition, **classifier_kwargs) + return torch.autograd.grad(log_prob.sum(), x_in)[0] + + def model_fn(x, t_continuous): + """ + The noise predicition model function that is used for DPM-Solver. + """ + if guidance_type == "uncond": + return noise_pred_fn(x, t_continuous) + elif guidance_type == "classifier": + assert classifier_fn is not None + t_input = get_model_input_time(t_continuous) + cond_grad = cond_grad_fn(x, t_input) + sigma_t = noise_schedule.marginal_std(t_continuous) + noise = noise_pred_fn(x, t_continuous) + return noise - guidance_scale * expand_dims(sigma_t, x.dim()) * cond_grad + elif guidance_type == "classifier-free": + if guidance_scale == 1.0 or unconditional_condition is None: + return noise_pred_fn(x, t_continuous, cond=condition) + x_in = torch.cat([x] * 2) + t_in = torch.cat([t_continuous] * 2) + c_in = torch.cat([unconditional_condition, condition]) + noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2) + return noise_uncond + guidance_scale * (noise - noise_uncond) + + assert model_type in ["noise", "x_start", "v", "score"] + assert guidance_type in ["uncond", "classifier", "classifier-free"] + return model_fn + + +class DPM_Solver: + def __init__( + self, + model_fn, + noise_schedule, + algorithm_type="dpmsolver++", + correcting_x0_fn=None, + correcting_xt_fn=None, + thresholding_max_val=1.0, + dynamic_thresholding_ratio=0.995, + ): + """Construct a DPM-Solver. + + We support both DPM-Solver (`algorithm_type="dpmsolver"`) and DPM-Solver++ (`algorithm_type="dpmsolver++"`). + + We also support the "dynamic thresholding" method in Imagen[1]. For pixel-space diffusion models, you + can set both `algorithm_type="dpmsolver++"` and `correcting_x0_fn="dynamic_thresholding"` to use the + dynamic thresholding. The "dynamic thresholding" can greatly improve the sample quality for pixel-space + DPMs with large guidance scales. Note that the thresholding method is **unsuitable** for latent-space + DPMs (such as stable-diffusion). + + To support advanced algorithms in image-to-image applications, we also support corrector functions for + both x0 and xt. + + Args: + model_fn: A noise prediction model function which accepts the continuous-time input (t in [epsilon, T]): + `` + def model_fn(x, t_continuous): + return noise + `` + The shape of `x` is `(batch_size, **shape)`, and the shape of `t_continuous` is `(batch_size,)`. + noise_schedule: A noise schedule object, such as NoiseScheduleVP. + algorithm_type: A `str`. Either "dpmsolver" or "dpmsolver++". + correcting_x0_fn: A `str` or a function with the following format: + ``` + def correcting_x0_fn(x0, t): + x0_new = ... + return x0_new + ``` + This function is to correct the outputs of the data prediction model at each sampling step. e.g., + ``` + x0_pred = data_pred_model(xt, t) + if correcting_x0_fn is not None: + x0_pred = correcting_x0_fn(x0_pred, t) + xt_1 = update(x0_pred, xt, t) + ``` + If `correcting_x0_fn="dynamic_thresholding"`, we use the dynamic thresholding proposed in Imagen[1]. + correcting_xt_fn: A function with the following format: + ``` + def correcting_xt_fn(xt, t, step): + x_new = ... + return x_new + ``` + This function is to correct the intermediate samples xt at each sampling step. e.g., + ``` + xt = ... + xt = correcting_xt_fn(xt, t, step) + ``` + thresholding_max_val: A `float`. The max value for thresholding. + Valid only when use `dpmsolver++` and `correcting_x0_fn="dynamic_thresholding"`. + dynamic_thresholding_ratio: A `float`. The ratio for dynamic thresholding (see Imagen[1] for details). + Valid only when use `dpmsolver++` and `correcting_x0_fn="dynamic_thresholding"`. + + [1] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, + Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models + with deep language understanding. arXiv preprint arXiv:2205.11487, 2022b. + """ + self.model = lambda x, t: model_fn(x, t.expand((x.shape[0]))) + self.noise_schedule = noise_schedule + assert algorithm_type in ["dpmsolver", "dpmsolver++"] + self.algorithm_type = algorithm_type + if correcting_x0_fn == "dynamic_thresholding": + self.correcting_x0_fn = self.dynamic_thresholding_fn + else: + self.correcting_x0_fn = correcting_x0_fn + self.correcting_xt_fn = correcting_xt_fn + self.dynamic_thresholding_ratio = dynamic_thresholding_ratio + self.thresholding_max_val = thresholding_max_val + + def dynamic_thresholding_fn(self, x0, t): + """ + The dynamic thresholding method. + """ + dims = x0.dim() + p = self.dynamic_thresholding_ratio + s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1) + s = expand_dims(torch.maximum(s, self.thresholding_max_val * torch.ones_like(s).to(s.device)), dims) + x0 = torch.clamp(x0, -s, s) / s + return x0 + + def noise_prediction_fn(self, x, t): + """ + Return the noise prediction model. + """ + return self.model(x, t) + + def data_prediction_fn(self, x, t): + """ + Return the data prediction model (with corrector). + """ + noise = self.noise_prediction_fn(x, t) + alpha_t, sigma_t = self.noise_schedule.marginal_alpha(t), self.noise_schedule.marginal_std(t) + x0 = (x - sigma_t * noise) / alpha_t + if self.correcting_x0_fn is not None: + x0 = self.correcting_x0_fn(x0, t) + return x0 + + def model_fn(self, x, t): + """ + Convert the model to the noise prediction model or the data prediction model. + """ + if self.algorithm_type == "dpmsolver++": + return self.data_prediction_fn(x, t) + else: + return self.noise_prediction_fn(x, t) + + def get_time_steps(self, skip_type, t_T, t_0, N, device): + """Compute the intermediate time steps for sampling. + + Args: + skip_type: A `str`. The type for the spacing of the time steps. We support three types: + - 'logSNR': uniform logSNR for the time steps. + - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.) + - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.) + t_T: A `float`. The starting time of the sampling (default is T). + t_0: A `float`. The ending time of the sampling (default is epsilon). + N: A `int`. The total number of the spacing of the time steps. + device: A torch device. + Returns: + A pytorch tensor of the time steps, with the shape (N + 1,). + """ + if skip_type == "logSNR": + lambda_T = self.noise_schedule.marginal_lambda(torch.tensor(t_T).to(device)) + lambda_0 = self.noise_schedule.marginal_lambda(torch.tensor(t_0).to(device)) + logSNR_steps = torch.linspace(lambda_T.cpu().item(), lambda_0.cpu().item(), N + 1).to(device) + return self.noise_schedule.inverse_lambda(logSNR_steps) + elif skip_type == "time_uniform": + return torch.linspace(t_T, t_0, N + 1).to(device) + elif skip_type == "time_quadratic": + t_order = 2 + return torch.linspace(t_T ** (1.0 / t_order), t_0 ** (1.0 / t_order), N + 1).pow(t_order).to(device) + else: + raise ValueError( + f"Unsupported skip_type {skip_type}, need to be 'logSNR' or 'time_uniform' or 'time_quadratic'" + ) + + def get_orders_and_timesteps_for_singlestep_solver(self, steps, order, skip_type, t_T, t_0, device): + """ + Get the order of each step for sampling by the singlestep DPM-Solver. + + We combine both DPM-Solver-1,2,3 to use all the function evaluations, which is named as "DPM-Solver-fast". + Given a fixed number of function evaluations by `steps`, the sampling procedure by DPM-Solver-fast is: + - If order == 1: + We take `steps` of DPM-Solver-1 (i.e. DDIM). + - If order == 2: + - Denote K = (steps // 2). We take K or (K + 1) intermediate time steps for sampling. + - If steps % 2 == 0, we use K steps of DPM-Solver-2. + - If steps % 2 == 1, we use K steps of DPM-Solver-2 and 1 step of DPM-Solver-1. + - If order == 3: + - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling. + - If steps % 3 == 0, we use (K - 2) steps of DPM-Solver-3, and 1 step of DPM-Solver-2 and 1 step of DPM-Solver-1. + - If steps % 3 == 1, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-1. + - If steps % 3 == 2, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-2. + + ============================================ + Args: + order: A `int`. The max order for the solver (2 or 3). + steps: A `int`. The total number of function evaluations (NFE). + skip_type: A `str`. The type for the spacing of the time steps. We support three types: + - 'logSNR': uniform logSNR for the time steps. + - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.) + - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.) + t_T: A `float`. The starting time of the sampling (default is T). + t_0: A `float`. The ending time of the sampling (default is epsilon). + device: A torch device. + Returns: + orders: A list of the solver order of each step. + """ + if order == 3: + K = steps // 3 + 1 + if steps % 3 == 0: + orders = [ + 3, + ] * ( + K - 2 + ) + [2, 1] + elif steps % 3 == 1: + orders = [ + 3, + ] * ( + K - 1 + ) + [1] + else: + orders = [ + 3, + ] * ( + K - 1 + ) + [2] + elif order == 2: + if steps % 2 == 0: + K = steps // 2 + orders = [ + 2, + ] * K + else: + K = steps // 2 + 1 + orders = [ + 2, + ] * ( + K - 1 + ) + [1] + elif order == 1: + K = 1 + orders = [ + 1, + ] * steps + else: + raise ValueError("'order' must be '1' or '2' or '3'.") + if skip_type == "logSNR": + # To reproduce the results in DPM-Solver paper + timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, K, device) + else: + timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, steps, device)[ + torch.cumsum( + torch.tensor( + [ + 0, + ] + + orders + ), + 0, + ).to(device) + ] + return timesteps_outer, orders + + def denoise_to_zero_fn(self, x, s): + """ + Denoise at the final step, which is equivalent to solve the ODE from lambda_s to infty by first-order discretization. + """ + return self.data_prediction_fn(x, s) + + def dpm_solver_first_update(self, x, s, t, model_s=None, return_intermediate=False): + """ + DPM-Solver-1 (equivalent to DDIM) from time `s` to time `t`. + + Args: + x: A pytorch tensor. The initial value at time `s`. + s: A pytorch tensor. The starting time, with the shape (1,). + t: A pytorch tensor. The ending time, with the shape (1,). + model_s: A pytorch tensor. The model function evaluated at time `s`. + If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it. + return_intermediate: A `bool`. If true, also return the model value at time `s`. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + ns = self.noise_schedule + x.dim() + lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t) + h = lambda_t - lambda_s + log_alpha_s, log_alpha_t = ns.marginal_log_mean_coeff(s), ns.marginal_log_mean_coeff(t) + sigma_s, sigma_t = ns.marginal_std(s), ns.marginal_std(t) + alpha_t = torch.exp(log_alpha_t) + + if self.algorithm_type == "dpmsolver++": + phi_1 = torch.expm1(-h) + if model_s is None: + model_s = self.model_fn(x, s) + x_t = sigma_t / sigma_s * x - alpha_t * phi_1 * model_s + else: + phi_1 = torch.expm1(h) + if model_s is None: + model_s = self.model_fn(x, s) + x_t = torch.exp(log_alpha_t - log_alpha_s) * x - (sigma_t * phi_1) * model_s + return (x_t, {"model_s": model_s}) if return_intermediate else x_t + + def singlestep_dpm_solver_second_update( + self, x, s, t, r1=0.5, model_s=None, return_intermediate=False, solver_type="dpmsolver" + ): + """ + Singlestep solver DPM-Solver-2 from time `s` to time `t`. + + Args: + x: A pytorch tensor. The initial value at time `s`. + s: A pytorch tensor. The starting time, with the shape (1,). + t: A pytorch tensor. The ending time, with the shape (1,). + r1: A `float`. The hyperparameter of the second-order solver. + model_s: A pytorch tensor. The model function evaluated at time `s`. + If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it. + return_intermediate: A `bool`. If true, also return the model value at time `s` and `s1` (the intermediate time). + solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpmsolver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if solver_type not in ["dpmsolver", "taylor"]: + raise ValueError(f"'solver_type' must be either 'dpmsolver' or 'taylor', got {solver_type}") + if r1 is None: + r1 = 0.5 + ns = self.noise_schedule + lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t) + h = lambda_t - lambda_s + lambda_s1 = lambda_s + r1 * h + s1 = ns.inverse_lambda(lambda_s1) + log_alpha_s, log_alpha_s1, log_alpha_t = ( + ns.marginal_log_mean_coeff(s), + ns.marginal_log_mean_coeff(s1), + ns.marginal_log_mean_coeff(t), + ) + sigma_s, sigma_s1, sigma_t = ns.marginal_std(s), ns.marginal_std(s1), ns.marginal_std(t) + alpha_s1, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_t) + + if self.algorithm_type == "dpmsolver++": + phi_11 = torch.expm1(-r1 * h) + phi_1 = torch.expm1(-h) + + if model_s is None: + model_s = self.model_fn(x, s) + x_s1 = (sigma_s1 / sigma_s) * x - (alpha_s1 * phi_11) * model_s + model_s1 = self.model_fn(x_s1, s1) + if solver_type == "dpmsolver": + x_t = ( + (sigma_t / sigma_s) * x + - (alpha_t * phi_1) * model_s + - (0.5 / r1) * (alpha_t * phi_1) * (model_s1 - model_s) + ) + elif solver_type == "taylor": + x_t = ( + (sigma_t / sigma_s) * x + - (alpha_t * phi_1) * model_s + + (1.0 / r1) * (alpha_t * (phi_1 / h + 1.0)) * (model_s1 - model_s) + ) + else: + phi_11 = torch.expm1(r1 * h) + phi_1 = torch.expm1(h) + + if model_s is None: + model_s = self.model_fn(x, s) + x_s1 = torch.exp(log_alpha_s1 - log_alpha_s) * x - (sigma_s1 * phi_11) * model_s + model_s1 = self.model_fn(x_s1, s1) + if solver_type == "dpmsolver": + x_t = ( + torch.exp(log_alpha_t - log_alpha_s) * x + - (sigma_t * phi_1) * model_s + - (0.5 / r1) * (sigma_t * phi_1) * (model_s1 - model_s) + ) + elif solver_type == "taylor": + x_t = ( + torch.exp(log_alpha_t - log_alpha_s) * x + - (sigma_t * phi_1) * model_s + - (1.0 / r1) * (sigma_t * (phi_1 / h - 1.0)) * (model_s1 - model_s) + ) + if return_intermediate: + return x_t, {"model_s": model_s, "model_s1": model_s1} + else: + return x_t + + def singlestep_dpm_solver_third_update( + self, + x, + s, + t, + r1=1.0 / 3.0, + r2=2.0 / 3.0, + model_s=None, + model_s1=None, + return_intermediate=False, + solver_type="dpmsolver", + ): + """ + Singlestep solver DPM-Solver-3 from time `s` to time `t`. + + Args: + x: A pytorch tensor. The initial value at time `s`. + s: A pytorch tensor. The starting time, with the shape (1,). + t: A pytorch tensor. The ending time, with the shape (1,). + r1: A `float`. The hyperparameter of the third-order solver. + r2: A `float`. The hyperparameter of the third-order solver. + model_s: A pytorch tensor. The model function evaluated at time `s`. + If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it. + model_s1: A pytorch tensor. The model function evaluated at time `s1` (the intermediate time given by `r1`). + If `model_s1` is None, we evaluate the model at `s1`; otherwise we directly use it. + return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times). + solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpmsolver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if solver_type not in ["dpmsolver", "taylor"]: + raise ValueError(f"'solver_type' must be either 'dpmsolver' or 'taylor', got {solver_type}") + if r1 is None: + r1 = 1.0 / 3.0 + if r2 is None: + r2 = 2.0 / 3.0 + ns = self.noise_schedule + lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t) + h = lambda_t - lambda_s + lambda_s1 = lambda_s + r1 * h + lambda_s2 = lambda_s + r2 * h + s1 = ns.inverse_lambda(lambda_s1) + s2 = ns.inverse_lambda(lambda_s2) + log_alpha_s, log_alpha_s1, log_alpha_s2, log_alpha_t = ( + ns.marginal_log_mean_coeff(s), + ns.marginal_log_mean_coeff(s1), + ns.marginal_log_mean_coeff(s2), + ns.marginal_log_mean_coeff(t), + ) + sigma_s, sigma_s1, sigma_s2, sigma_t = ( + ns.marginal_std(s), + ns.marginal_std(s1), + ns.marginal_std(s2), + ns.marginal_std(t), + ) + alpha_s1, alpha_s2, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_s2), torch.exp(log_alpha_t) + + if self.algorithm_type == "dpmsolver++": + phi_11 = torch.expm1(-r1 * h) + phi_12 = torch.expm1(-r2 * h) + phi_1 = torch.expm1(-h) + phi_22 = torch.expm1(-r2 * h) / (r2 * h) + 1.0 + phi_2 = phi_1 / h + 1.0 + phi_3 = phi_2 / h - 0.5 + + if model_s is None: + model_s = self.model_fn(x, s) + if model_s1 is None: + x_s1 = (sigma_s1 / sigma_s) * x - (alpha_s1 * phi_11) * model_s + model_s1 = self.model_fn(x_s1, s1) + x_s2 = ( + (sigma_s2 / sigma_s) * x + - (alpha_s2 * phi_12) * model_s + + r2 / r1 * (alpha_s2 * phi_22) * (model_s1 - model_s) + ) + model_s2 = self.model_fn(x_s2, s2) + if solver_type == "dpmsolver": + x_t = ( + (sigma_t / sigma_s) * x + - (alpha_t * phi_1) * model_s + + (1.0 / r2) * (alpha_t * phi_2) * (model_s2 - model_s) + ) + elif solver_type == "taylor": + D1_0 = (1.0 / r1) * (model_s1 - model_s) + D1_1 = (1.0 / r2) * (model_s2 - model_s) + D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1) + D2 = 2.0 * (D1_1 - D1_0) / (r2 - r1) + x_t = ( + (sigma_t / sigma_s) * x + - (alpha_t * phi_1) * model_s + + (alpha_t * phi_2) * D1 + - (alpha_t * phi_3) * D2 + ) + else: + phi_11 = torch.expm1(r1 * h) + phi_12 = torch.expm1(r2 * h) + phi_1 = torch.expm1(h) + phi_22 = torch.expm1(r2 * h) / (r2 * h) - 1.0 + phi_2 = phi_1 / h - 1.0 + phi_3 = phi_2 / h - 0.5 + + if model_s is None: + model_s = self.model_fn(x, s) + if model_s1 is None: + x_s1 = (torch.exp(log_alpha_s1 - log_alpha_s)) * x - (sigma_s1 * phi_11) * model_s + model_s1 = self.model_fn(x_s1, s1) + x_s2 = ( + (torch.exp(log_alpha_s2 - log_alpha_s)) * x + - (sigma_s2 * phi_12) * model_s + - r2 / r1 * (sigma_s2 * phi_22) * (model_s1 - model_s) + ) + model_s2 = self.model_fn(x_s2, s2) + if solver_type == "dpmsolver": + x_t = ( + (torch.exp(log_alpha_t - log_alpha_s)) * x + - (sigma_t * phi_1) * model_s + - (1.0 / r2) * (sigma_t * phi_2) * (model_s2 - model_s) + ) + elif solver_type == "taylor": + D1_0 = (1.0 / r1) * (model_s1 - model_s) + D1_1 = (1.0 / r2) * (model_s2 - model_s) + D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1) + D2 = 2.0 * (D1_1 - D1_0) / (r2 - r1) + x_t = ( + (torch.exp(log_alpha_t - log_alpha_s)) * x + - (sigma_t * phi_1) * model_s + - (sigma_t * phi_2) * D1 + - (sigma_t * phi_3) * D2 + ) + + if return_intermediate: + return x_t, {"model_s": model_s, "model_s1": model_s1, "model_s2": model_s2} + else: + return x_t + + def multistep_dpm_solver_second_update(self, x, model_prev_list, t_prev_list, t, solver_type="dpmsolver"): + """ + Multistep solver DPM-Solver-2 from time `t_prev_list[-1]` to time `t`. + + Args: + x: A pytorch tensor. The initial value at time `s`. + model_prev_list: A list of pytorch tensor. The previous computed model values. + t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (1,) + t: A pytorch tensor. The ending time, with the shape (1,). + solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpmsolver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if solver_type not in ["dpmsolver", "taylor"]: + raise ValueError(f"'solver_type' must be either 'dpmsolver' or 'taylor', got {solver_type}") + ns = self.noise_schedule + model_prev_1, model_prev_0 = model_prev_list[-2], model_prev_list[-1] + t_prev_1, t_prev_0 = t_prev_list[-2], t_prev_list[-1] + lambda_prev_1, lambda_prev_0, lambda_t = ( + ns.marginal_lambda(t_prev_1), + ns.marginal_lambda(t_prev_0), + ns.marginal_lambda(t), + ) + log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t) + sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t) + alpha_t = torch.exp(log_alpha_t) + + h_0 = lambda_prev_0 - lambda_prev_1 + h = lambda_t - lambda_prev_0 + r0 = h_0 / h + D1_0 = (1.0 / r0) * (model_prev_0 - model_prev_1) + if self.algorithm_type == "dpmsolver++": + phi_1 = torch.expm1(-h) + if solver_type == "dpmsolver": + x_t = (sigma_t / sigma_prev_0) * x - (alpha_t * phi_1) * model_prev_0 - 0.5 * (alpha_t * phi_1) * D1_0 + elif solver_type == "taylor": + x_t = ( + (sigma_t / sigma_prev_0) * x + - (alpha_t * phi_1) * model_prev_0 + + (alpha_t * (phi_1 / h + 1.0)) * D1_0 + ) + else: + phi_1 = torch.expm1(h) + if solver_type == "dpmsolver": + x_t = ( + (torch.exp(log_alpha_t - log_alpha_prev_0)) * x + - (sigma_t * phi_1) * model_prev_0 + - 0.5 * (sigma_t * phi_1) * D1_0 + ) + elif solver_type == "taylor": + x_t = ( + (torch.exp(log_alpha_t - log_alpha_prev_0)) * x + - (sigma_t * phi_1) * model_prev_0 + - (sigma_t * (phi_1 / h - 1.0)) * D1_0 + ) + return x_t + + def multistep_dpm_solver_third_update(self, x, model_prev_list, t_prev_list, t, solver_type="dpmsolver"): + """ + Multistep solver DPM-Solver-3 from time `t_prev_list[-1]` to time `t`. + + Args: + x: A pytorch tensor. The initial value at time `s`. + model_prev_list: A list of pytorch tensor. The previous computed model values. + t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (1,) + t: A pytorch tensor. The ending time, with the shape (1,). + solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpmsolver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + ns = self.noise_schedule + model_prev_2, model_prev_1, model_prev_0 = model_prev_list + t_prev_2, t_prev_1, t_prev_0 = t_prev_list + lambda_prev_2, lambda_prev_1, lambda_prev_0, lambda_t = ( + ns.marginal_lambda(t_prev_2), + ns.marginal_lambda(t_prev_1), + ns.marginal_lambda(t_prev_0), + ns.marginal_lambda(t), + ) + log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t) + sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t) + alpha_t = torch.exp(log_alpha_t) + + h_1 = lambda_prev_1 - lambda_prev_2 + h_0 = lambda_prev_0 - lambda_prev_1 + h = lambda_t - lambda_prev_0 + r0, r1 = h_0 / h, h_1 / h + D1_0 = (1.0 / r0) * (model_prev_0 - model_prev_1) + D1_1 = (1.0 / r1) * (model_prev_1 - model_prev_2) + D1 = D1_0 + (r0 / (r0 + r1)) * (D1_0 - D1_1) + D2 = (1.0 / (r0 + r1)) * (D1_0 - D1_1) + if self.algorithm_type == "dpmsolver++": + phi_1 = torch.expm1(-h) + phi_2 = phi_1 / h + 1.0 + phi_3 = phi_2 / h - 0.5 + return ( + (sigma_t / sigma_prev_0) * x + - (alpha_t * phi_1) * model_prev_0 + + (alpha_t * phi_2) * D1 + - (alpha_t * phi_3) * D2 + ) + else: + phi_1 = torch.expm1(h) + phi_2 = phi_1 / h - 1.0 + phi_3 = phi_2 / h - 0.5 + return ( + (torch.exp(log_alpha_t - log_alpha_prev_0)) * x + - (sigma_t * phi_1) * model_prev_0 + - (sigma_t * phi_2) * D1 + - (sigma_t * phi_3) * D2 + ) + + def singlestep_dpm_solver_update( + self, x, s, t, order, return_intermediate=False, solver_type="dpmsolver", r1=None, r2=None + ): + """ + Singlestep DPM-Solver with the order `order` from time `s` to time `t`. + + Args: + x: A pytorch tensor. The initial value at time `s`. + s: A pytorch tensor. The starting time, with the shape (1,). + t: A pytorch tensor. The ending time, with the shape (1,). + order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3. + return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times). + solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpmsolver' type. + r1: A `float`. The hyperparameter of the second-order or third-order solver. + r2: A `float`. The hyperparameter of the third-order solver. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if order == 1: + return self.dpm_solver_first_update(x, s, t, return_intermediate=return_intermediate) + elif order == 2: + return self.singlestep_dpm_solver_second_update( + x, s, t, return_intermediate=return_intermediate, solver_type=solver_type, r1=r1 + ) + elif order == 3: + return self.singlestep_dpm_solver_third_update( + x, s, t, return_intermediate=return_intermediate, solver_type=solver_type, r1=r1, r2=r2 + ) + else: + raise ValueError(f"Solver order must be 1 or 2 or 3, got {order}") + + def multistep_dpm_solver_update(self, x, model_prev_list, t_prev_list, t, order, solver_type="dpmsolver"): + """ + Multistep DPM-Solver with the order `order` from time `t_prev_list[-1]` to time `t`. + + Args: + x: A pytorch tensor. The initial value at time `s`. + model_prev_list: A list of pytorch tensor. The previous computed model values. + t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (1,) + t: A pytorch tensor. The ending time, with the shape (1,). + order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3. + solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpmsolver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if order == 1: + return self.dpm_solver_first_update(x, t_prev_list[-1], t, model_s=model_prev_list[-1]) + elif order == 2: + return self.multistep_dpm_solver_second_update(x, model_prev_list, t_prev_list, t, solver_type=solver_type) + elif order == 3: + return self.multistep_dpm_solver_third_update(x, model_prev_list, t_prev_list, t, solver_type=solver_type) + else: + raise ValueError(f"Solver order must be 1 or 2 or 3, got {order}") + + def dpm_solver_adaptive( + self, x, order, t_T, t_0, h_init=0.05, atol=0.0078, rtol=0.05, theta=0.9, t_err=1e-5, solver_type="dpmsolver" + ): + """ + The adaptive step size solver based on singlestep DPM-Solver. + + Args: + x: A pytorch tensor. The initial value at time `t_T`. + order: A `int`. The (higher) order of the solver. We only support order == 2 or 3. + t_T: A `float`. The starting time of the sampling (default is T). + t_0: A `float`. The ending time of the sampling (default is epsilon). + h_init: A `float`. The initial step size (for logSNR). + atol: A `float`. The absolute tolerance of the solver. For image data, the default setting is 0.0078, followed [1]. + rtol: A `float`. The relative tolerance of the solver. The default setting is 0.05. + theta: A `float`. The safety hyperparameter for adapting the step size. The default setting is 0.9, followed [1]. + t_err: A `float`. The tolerance for the time. We solve the diffusion ODE until the absolute error between the + current time and `t_0` is less than `t_err`. The default setting is 1e-5. + solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpmsolver' type. + Returns: + x_0: A pytorch tensor. The approximated solution at time `t_0`. + + [1] A. Jolicoeur-Martineau, K. Li, R. Piché-Taillefer, T. Kachman, and I. Mitliagkas, "Gotta go fast when generating data with score-based models," arXiv preprint arXiv:2105.14080, 2021. + """ + ns = self.noise_schedule + s = t_T * torch.ones((1,)).to(x) + lambda_s = ns.marginal_lambda(s) + lambda_0 = ns.marginal_lambda(t_0 * torch.ones_like(s).to(x)) + h = h_init * torch.ones_like(s).to(x) + x_prev = x + nfe = 0 + if order == 2: + r1 = 0.5 + lower_update = lambda x, s, t: self.dpm_solver_first_update(x, s, t, return_intermediate=True) + higher_update = lambda x, s, t, **kwargs: self.singlestep_dpm_solver_second_update( + x, s, t, r1=r1, solver_type=solver_type, **kwargs + ) + elif order == 3: + r1, r2 = 1.0 / 3.0, 2.0 / 3.0 + lower_update = lambda x, s, t: self.singlestep_dpm_solver_second_update( + x, s, t, r1=r1, return_intermediate=True, solver_type=solver_type + ) + higher_update = lambda x, s, t, **kwargs: self.singlestep_dpm_solver_third_update( + x, s, t, r1=r1, r2=r2, solver_type=solver_type, **kwargs + ) + else: + raise ValueError(f"For adaptive step size solver, order must be 2 or 3, got {order}") + while torch.abs((s - t_0)).mean() > t_err: + t = ns.inverse_lambda(lambda_s + h) + x_lower, lower_noise_kwargs = lower_update(x, s, t) + x_higher = higher_update(x, s, t, **lower_noise_kwargs) + delta = torch.max(torch.ones_like(x).to(x) * atol, rtol * torch.max(torch.abs(x_lower), torch.abs(x_prev))) + norm_fn = lambda v: torch.sqrt(torch.square(v.reshape((v.shape[0], -1))).mean(dim=-1, keepdim=True)) + E = norm_fn((x_higher - x_lower) / delta).max() + if torch.all(E <= 1.0): + x = x_higher + s = t + x_prev = x_lower + lambda_s = ns.marginal_lambda(s) + h = torch.min(theta * h * torch.float_power(E, -1.0 / order).float(), lambda_0 - lambda_s) + nfe += order + print("adaptive solver nfe", nfe) + return x + + def add_noise(self, x, t, noise=None): + """ + Compute the noised input xt = alpha_t * x + sigma_t * noise. + + Args: + x: A `torch.Tensor` with shape `(batch_size, *shape)`. + t: A `torch.Tensor` with shape `(t_size,)`. + Returns: + xt with shape `(t_size, batch_size, *shape)`. + """ + alpha_t, sigma_t = self.noise_schedule.marginal_alpha(t), self.noise_schedule.marginal_std(t) + if noise is None: + noise = torch.randn((t.shape[0], *x.shape), device=x.device) + x = x.reshape((-1, *x.shape)) + xt = expand_dims(alpha_t, x.dim()) * x + expand_dims(sigma_t, x.dim()) * noise + return xt.squeeze(0) if t.shape[0] == 1 else xt + + def inverse( + self, + x, + steps=20, + t_start=None, + t_end=None, + order=2, + skip_type="time_uniform", + method="multistep", + lower_order_final=True, + denoise_to_zero=False, + solver_type="dpmsolver", + atol=0.0078, + rtol=0.05, + return_intermediate=False, + ): + """ + Inverse the sample `x` from time `t_start` to `t_end` by DPM-Solver. + For discrete-time DPMs, we use `t_start=1/N`, where `N` is the total time steps during training. + """ + t_0 = 1.0 / self.noise_schedule.total_N if t_start is None else t_start + t_T = self.noise_schedule.T if t_end is None else t_end + assert ( + t_0 > 0 and t_T > 0 + ), "Time range needs to be greater than 0. For discrete-time DPMs, it needs to be in [1 / N, 1], where N is the length of betas array" + return self.sample( + x, + steps=steps, + t_start=t_0, + t_end=t_T, + order=order, + skip_type=skip_type, + method=method, + lower_order_final=lower_order_final, + denoise_to_zero=denoise_to_zero, + solver_type=solver_type, + atol=atol, + rtol=rtol, + return_intermediate=return_intermediate, + ) + + def sample( + self, + x, + steps=20, + t_start=None, + t_end=None, + order=2, + skip_type="time_uniform", + method="multistep", + lower_order_final=True, + denoise_to_zero=False, + solver_type="dpmsolver", + atol=0.0078, + rtol=0.05, + return_intermediate=False, + ): + """ + Compute the sample at time `t_end` by DPM-Solver, given the initial `x` at time `t_start`. + + ===================================================== + + We support the following algorithms for both noise prediction model and data prediction model: + - 'singlestep': + Singlestep DPM-Solver (i.e. "DPM-Solver-fast" in the paper), which combines different orders of singlestep DPM-Solver. + We combine all the singlestep solvers with order <= `order` to use up all the function evaluations (steps). + The total number of function evaluations (NFE) == `steps`. + Given a fixed NFE == `steps`, the sampling procedure is: + - If `order` == 1: + - Denote K = steps. We use K steps of DPM-Solver-1 (i.e. DDIM). + - If `order` == 2: + - Denote K = (steps // 2) + (steps % 2). We take K intermediate time steps for sampling. + - If steps % 2 == 0, we use K steps of singlestep DPM-Solver-2. + - If steps % 2 == 1, we use (K - 1) steps of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1. + - If `order` == 3: + - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling. + - If steps % 3 == 0, we use (K - 2) steps of singlestep DPM-Solver-3, and 1 step of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1. + - If steps % 3 == 1, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of DPM-Solver-1. + - If steps % 3 == 2, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of singlestep DPM-Solver-2. + - 'multistep': + Multistep DPM-Solver with the order of `order`. The total number of function evaluations (NFE) == `steps`. + We initialize the first `order` values by lower order multistep solvers. + Given a fixed NFE == `steps`, the sampling procedure is: + Denote K = steps. + - If `order` == 1: + - We use K steps of DPM-Solver-1 (i.e. DDIM). + - If `order` == 2: + - We firstly use 1 step of DPM-Solver-1, then use (K - 1) step of multistep DPM-Solver-2. + - If `order` == 3: + - We firstly use 1 step of DPM-Solver-1, then 1 step of multistep DPM-Solver-2, then (K - 2) step of multistep DPM-Solver-3. + - 'singlestep_fixed': + Fixed order singlestep DPM-Solver (i.e. DPM-Solver-1 or singlestep DPM-Solver-2 or singlestep DPM-Solver-3). + We use singlestep DPM-Solver-`order` for `order`=1 or 2 or 3, with total [`steps` // `order`] * `order` NFE. + - 'adaptive': + Adaptive step size DPM-Solver (i.e. "DPM-Solver-12" and "DPM-Solver-23" in the paper). + We ignore `steps` and use adaptive step size DPM-Solver with a higher order of `order`. + You can adjust the absolute tolerance `atol` and the relative tolerance `rtol` to balance the computatation costs + (NFE) and the sample quality. + - If `order` == 2, we use DPM-Solver-12 which combines DPM-Solver-1 and singlestep DPM-Solver-2. + - If `order` == 3, we use DPM-Solver-23 which combines singlestep DPM-Solver-2 and singlestep DPM-Solver-3. + + ===================================================== + + Some advices for choosing the algorithm: + - For **unconditional sampling** or **guided sampling with small guidance scale** by DPMs: + Use singlestep DPM-Solver or DPM-Solver++ ("DPM-Solver-fast" in the paper) with `order = 3`. + e.g., DPM-Solver: + >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, algorithm_type="dpmsolver") + >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=3, + skip_type='time_uniform', method='singlestep') + e.g., DPM-Solver++: + >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, algorithm_type="dpmsolver++") + >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=3, + skip_type='time_uniform', method='singlestep') + - For **guided sampling with large guidance scale** by DPMs: + Use multistep DPM-Solver with `algorithm_type="dpmsolver++"` and `order = 2`. + e.g. + >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, algorithm_type="dpmsolver++") + >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=2, + skip_type='time_uniform', method='multistep') + + We support three types of `skip_type`: + - 'logSNR': uniform logSNR for the time steps. **Recommended for low-resolutional images** + - 'time_uniform': uniform time for the time steps. **Recommended for high-resolutional images**. + - 'time_quadratic': quadratic time for the time steps. + + ===================================================== + Args: + x: A pytorch tensor. The initial value at time `t_start` + e.g. if `t_start` == T, then `x` is a sample from the standard normal distribution. + steps: A `int`. The total number of function evaluations (NFE). + t_start: A `float`. The starting time of the sampling. + If `T` is None, we use self.noise_schedule.T (default is 1.0). + t_end: A `float`. The ending time of the sampling. + If `t_end` is None, we use 1. / self.noise_schedule.total_N. + e.g. if total_N == 1000, we have `t_end` == 1e-3. + For discrete-time DPMs: + - We recommend `t_end` == 1. / self.noise_schedule.total_N. + For continuous-time DPMs: + - We recommend `t_end` == 1e-3 when `steps` <= 15; and `t_end` == 1e-4 when `steps` > 15. + order: A `int`. The order of DPM-Solver. + skip_type: A `str`. The type for the spacing of the time steps. 'time_uniform' or 'logSNR' or 'time_quadratic'. + method: A `str`. The method for sampling. 'singlestep' or 'multistep' or 'singlestep_fixed' or 'adaptive'. + denoise_to_zero: A `bool`. Whether to denoise to time 0 at the final step. + Default is `False`. If `denoise_to_zero` is `True`, the total NFE is (`steps` + 1). + + This trick is firstly proposed by DDPM (https://arxiv.org/abs/2006.11239) and + score_sde (https://arxiv.org/abs/2011.13456). Such trick can improve the FID + for diffusion models sampling by diffusion SDEs for low-resolutional images + (such as CIFAR-10). However, we observed that such trick does not matter for + high-resolutional images. As it needs an additional NFE, we do not recommend + it for high-resolutional images. + lower_order_final: A `bool`. Whether to use lower order solvers at the final steps. + Only valid for `method=multistep` and `steps < 15`. We empirically find that + this trick is a key to stabilizing the sampling by DPM-Solver with very few steps + (especially for steps <= 10). So we recommend to set it to be `True`. + solver_type: A `str`. The taylor expansion type for the solver. `dpmsolver` or `taylor`. We recommend `dpmsolver`. + atol: A `float`. The absolute tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'. + rtol: A `float`. The relative tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'. + return_intermediate: A `bool`. Whether to save the xt at each step. + When set to `True`, method returns a tuple (x0, intermediates); when set to False, method returns only x0. + Returns: + x_end: A pytorch tensor. The approximated solution at time `t_end`. + + """ + t_0 = 1.0 / self.noise_schedule.total_N if t_end is None else t_end + t_T = self.noise_schedule.T if t_start is None else t_start + assert ( + t_0 > 0 and t_T > 0 + ), "Time range needs to be greater than 0. For discrete-time DPMs, it needs to be in [1 / N, 1], where N is the length of betas array" + if return_intermediate: + assert method in [ + "multistep", + "singlestep", + "singlestep_fixed", + ], "Cannot use adaptive solver when saving intermediate values" + if self.correcting_xt_fn is not None: + assert method in [ + "multistep", + "singlestep", + "singlestep_fixed", + ], "Cannot use adaptive solver when correcting_xt_fn is not None" + device = x.device + intermediates = [] + with torch.no_grad(): + if method == "adaptive": + x = self.dpm_solver_adaptive( + x, order=order, t_T=t_T, t_0=t_0, atol=atol, rtol=rtol, solver_type=solver_type + ) + elif method == "multistep": + assert steps >= order + timesteps = self.get_time_steps(skip_type=skip_type, t_T=t_T, t_0=t_0, N=steps, device=device) + assert timesteps.shape[0] - 1 == steps + # Init the initial values. + step = 0 + t = timesteps[step] + t_prev_list = [t] + model_prev_list = [self.model_fn(x, t)] + if self.correcting_xt_fn is not None: + x = self.correcting_xt_fn(x, t, step) + if return_intermediate: + intermediates.append(x) + # Init the first `order` values by lower order multistep DPM-Solver. + for step in range(1, order): + t = timesteps[step] + x = self.multistep_dpm_solver_update( + x, model_prev_list, t_prev_list, t, step, solver_type=solver_type + ) + if self.correcting_xt_fn is not None: + x = self.correcting_xt_fn(x, t, step) + if return_intermediate: + intermediates.append(x) + t_prev_list.append(t) + model_prev_list.append(self.model_fn(x, t)) + # Compute the remaining values by `order`-th order multistep DPM-Solver. + for step in tqdm(range(order, steps + 1)): + t = timesteps[step] + # We only use lower order for steps < 10 + if lower_order_final and steps < 10: + step_order = min(order, steps + 1 - step) + else: + step_order = order + x = self.multistep_dpm_solver_update( + x, model_prev_list, t_prev_list, t, step_order, solver_type=solver_type + ) + if self.correcting_xt_fn is not None: + x = self.correcting_xt_fn(x, t, step) + if return_intermediate: + intermediates.append(x) + for i in range(order - 1): + t_prev_list[i] = t_prev_list[i + 1] + model_prev_list[i] = model_prev_list[i + 1] + t_prev_list[-1] = t + # We do not need to evaluate the final model value. + if step < steps: + model_prev_list[-1] = self.model_fn(x, t) + elif method in ["singlestep", "singlestep_fixed"]: + if method == "singlestep": + timesteps_outer, orders = self.get_orders_and_timesteps_for_singlestep_solver( + steps=steps, order=order, skip_type=skip_type, t_T=t_T, t_0=t_0, device=device + ) + elif method == "singlestep_fixed": + K = steps // order + orders = [ + order, + ] * K + timesteps_outer = self.get_time_steps(skip_type=skip_type, t_T=t_T, t_0=t_0, N=K, device=device) + for step, order in enumerate(orders): + s, t = timesteps_outer[step], timesteps_outer[step + 1] + timesteps_inner = self.get_time_steps( + skip_type=skip_type, t_T=s.item(), t_0=t.item(), N=order, device=device + ) + lambda_inner = self.noise_schedule.marginal_lambda(timesteps_inner) + h = lambda_inner[-1] - lambda_inner[0] + r1 = None if order <= 1 else (lambda_inner[1] - lambda_inner[0]) / h + r2 = None if order <= 2 else (lambda_inner[2] - lambda_inner[0]) / h + x = self.singlestep_dpm_solver_update(x, s, t, order, solver_type=solver_type, r1=r1, r2=r2) + if self.correcting_xt_fn is not None: + x = self.correcting_xt_fn(x, t, step) + if return_intermediate: + intermediates.append(x) + else: + raise ValueError(f"Got wrong method {method}") + if denoise_to_zero: + t = torch.ones((1,)).to(device) * t_0 + x = self.denoise_to_zero_fn(x, t) + if self.correcting_xt_fn is not None: + x = self.correcting_xt_fn(x, t, step + 1) + if return_intermediate: + intermediates.append(x) + return (x, intermediates) if return_intermediate else x + + +############################################################# +# other utility functions +############################################################# + + +def interpolate_fn(x, xp, yp): + """ + A piecewise linear function y = f(x), using xp and yp as keypoints. + We implement f(x) in a differentiable way (i.e. applicable for autograd). + The function f(x) is well-defined for all x-axis. (For x beyond the bounds of xp, we use the outmost points of xp to define the linear function.) + + Args: + x: PyTorch tensor with shape [N, C], where N is the batch size, C is the number of channels (we use C = 1 for DPM-Solver). + xp: PyTorch tensor with shape [C, K], where K is the number of keypoints. + yp: PyTorch tensor with shape [C, K]. + Returns: + The function values f(x), with shape [N, C]. + """ + N, K = x.shape[0], xp.shape[1] + all_x = torch.cat([x.unsqueeze(2), xp.unsqueeze(0).repeat((N, 1, 1))], dim=2) + sorted_all_x, x_indices = torch.sort(all_x, dim=2) + x_idx = torch.argmin(x_indices, dim=2) + cand_start_idx = x_idx - 1 + start_idx = torch.where( + torch.eq(x_idx, 0), + torch.tensor(1, device=x.device), + torch.where( + torch.eq(x_idx, K), + torch.tensor(K - 2, device=x.device), + cand_start_idx, + ), + ) + end_idx = torch.where(torch.eq(start_idx, cand_start_idx), start_idx + 2, start_idx + 1) + start_x = torch.gather(sorted_all_x, dim=2, index=start_idx.unsqueeze(2)).squeeze(2) + end_x = torch.gather(sorted_all_x, dim=2, index=end_idx.unsqueeze(2)).squeeze(2) + start_idx2 = torch.where( + torch.eq(x_idx, 0), + torch.tensor(0, device=x.device), + torch.where( + torch.eq(x_idx, K), + torch.tensor(K - 2, device=x.device), + cand_start_idx, + ), + ) + y_positions_expanded = yp.unsqueeze(0).expand(N, -1, -1) + start_y = torch.gather(y_positions_expanded, dim=2, index=start_idx2.unsqueeze(2)).squeeze(2) + end_y = torch.gather(y_positions_expanded, dim=2, index=(start_idx2 + 1).unsqueeze(2)).squeeze(2) + return start_y + (x - start_x) * (end_y - start_y) / (end_x - start_x) + + +def expand_dims(v, dims): + """ + Expand the tensor `v` to the dim `dims`. + + Args: + `v`: a PyTorch tensor with shape [N]. + `dim`: a `int`. + Returns: + a PyTorch tensor with shape [N, 1, 1, ..., 1] and the total dimension is `dims`. + """ + return v[(...,) + (None,) * (dims - 1)] + + +def DPMS( + model, + condition, + uncondition, + cfg_scale, + model_type="noise", + noise_schedule="linear", + guidance_type="classifier-free", + model_kwargs=None, + diffusion_steps=1000, +): + if model_kwargs is None: + model_kwargs = {} + betas = torch.tensor(get_named_beta_schedule(noise_schedule, diffusion_steps)) + + ## 1. Define the noise schedule. + noise_schedule = NoiseScheduleVP(schedule="discrete", betas=betas) + + ## 2. Convert your discrete-time `model` to the continuous-time + ## noise prediction model. Here is an example for a diffusion model + ## `model` with the noise prediction type ("noise") . + model_fn = model_wrapper( + model, + noise_schedule, + model_type=model_type, + model_kwargs=model_kwargs, + guidance_type=guidance_type, + condition=condition, + unconditional_condition=uncondition, + guidance_scale=cfg_scale, + ) + ## 3. Define dpm-solver and sample by multistep DPM-Solver. + return DPM_Solver(model_fn, noise_schedule, algorithm_type="dpmsolver++") diff --git a/opensora/schedulers/iddpm/__init__.py b/opensora/schedulers/iddpm/__init__.py new file mode 100644 index 00000000..b9806ad3 --- /dev/null +++ b/opensora/schedulers/iddpm/__init__.py @@ -0,0 +1,95 @@ +from functools import partial + +import torch + +from opensora.registry import SCHEDULERS + +from . import gaussian_diffusion as gd +from .respace import SpacedDiffusion, space_timesteps + + +@SCHEDULERS.register_module("iddpm") +class IDDPM(SpacedDiffusion): + def __init__( + self, + num_sampling_steps=None, + timestep_respacing=None, + noise_schedule="linear", + use_kl=False, + sigma_small=False, + predict_xstart=False, + learn_sigma=True, + rescale_learned_sigmas=False, + diffusion_steps=1000, + cfg_scale=4.0, + ): + betas = gd.get_named_beta_schedule(noise_schedule, diffusion_steps) + if use_kl: + loss_type = gd.LossType.RESCALED_KL + elif rescale_learned_sigmas: + loss_type = gd.LossType.RESCALED_MSE + else: + loss_type = gd.LossType.MSE + if num_sampling_steps is not None: + assert timestep_respacing is None + timestep_respacing = str(num_sampling_steps) + if timestep_respacing is None or timestep_respacing == "": + timestep_respacing = [diffusion_steps] + super().__init__( + use_timesteps=space_timesteps(diffusion_steps, timestep_respacing), + betas=betas, + model_mean_type=(gd.ModelMeanType.EPSILON if not predict_xstart else gd.ModelMeanType.START_X), + model_var_type=( + (gd.ModelVarType.FIXED_LARGE if not sigma_small else gd.ModelVarType.FIXED_SMALL) + if not learn_sigma + else gd.ModelVarType.LEARNED_RANGE + ), + loss_type=loss_type, + # rescale_timesteps=rescale_timesteps, + ) + + self.cfg_scale = cfg_scale + + def sample( + self, + model, + text_encoder, + z_size, + prompts, + device, + additional_args=None, + ): + n = len(prompts) + z = torch.randn(n, *z_size, device=device) + z = torch.cat([z, z], 0) + model_args = text_encoder.encode(prompts) + y_null = text_encoder.null(n) + model_args["y"] = torch.cat([model_args["y"], y_null], 0) + if additional_args is not None: + model_args.update(additional_args) + + forward = partial(forward_with_cfg, model, cfg_scale=self.cfg_scale) + samples = self.p_sample_loop( + forward, + z.shape, + z, + clip_denoised=False, + model_kwargs=model_args, + progress=True, + device=device, + ) + samples, _ = samples.chunk(2, dim=0) + return samples + + +def forward_with_cfg(model, x, timestep, y, cfg_scale, **kwargs): + # https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb + half = x[: len(x) // 2] + combined = torch.cat([half, half], dim=0) + model_out = model.forward(combined, timestep, y, **kwargs) + model_out = model_out["x"] if isinstance(model_out, dict) else model_out + eps, rest = model_out[:, :3], model_out[:, 3:] + cond_eps, uncond_eps = torch.split(eps, len(eps) // 2, dim=0) + half_eps = uncond_eps + cfg_scale * (cond_eps - uncond_eps) + eps = torch.cat([half_eps, half_eps], dim=0) + return torch.cat([eps, rest], dim=1) diff --git a/open_sora/diffusion/diffusion_utils.py b/opensora/schedulers/iddpm/diffusion_utils.py similarity index 82% rename from open_sora/diffusion/diffusion_utils.py rename to opensora/schedulers/iddpm/diffusion_utils.py index 056471c0..c097ac59 100644 --- a/open_sora/diffusion/diffusion_utils.py +++ b/opensora/schedulers/iddpm/diffusion_utils.py @@ -1,7 +1,15 @@ -# Modified from OpenAI's diffusion repos -# GLIDE: https://github.com/openai/glide-text2im/blob/main/glide_text2im/gaussian_diffusion.py -# ADM: https://github.com/openai/guided-diffusion/blob/main/guided_diffusion -# IDDPM: https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py +# Adapted from DiT + +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. +# -------------------------------------------------------- +# References: +# DiT: https://github.com/facebookresearch/DiT/tree/main +# GLIDE: https://github.com/openai/glide-text2im/blob/main/glide_text2im/gaussian_diffusion.py +# ADM: https://github.com/openai/guided-diffusion/blob/main/guided_diffusion +# IDDPM: https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py +# -------------------------------------------------------- + import numpy as np import torch as th diff --git a/open_sora/diffusion/gaussian_diffusion.py b/opensora/schedulers/iddpm/gaussian_diffusion.py similarity index 95% rename from open_sora/diffusion/gaussian_diffusion.py rename to opensora/schedulers/iddpm/gaussian_diffusion.py index fe94622f..4a74592c 100644 --- a/open_sora/diffusion/gaussian_diffusion.py +++ b/opensora/schedulers/iddpm/gaussian_diffusion.py @@ -1,8 +1,14 @@ -# Modified from OpenAI's diffusion repos -# GLIDE: https://github.com/openai/glide-text2im/blob/main/glide_text2im/gaussian_diffusion.py -# ADM: https://github.com/openai/guided-diffusion/blob/main/guided_diffusion -# IDDPM: https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py - +# Adapted from DiT + +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. +# -------------------------------------------------------- +# References: +# DiT: https://github.com/facebookresearch/DiT/tree/main +# GLIDE: https://github.com/openai/glide-text2im/blob/main/glide_text2im/gaussian_diffusion.py +# ADM: https://github.com/openai/guided-diffusion/blob/main/guided_diffusion +# IDDPM: https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py +# -------------------------------------------------------- import enum import math @@ -13,12 +19,10 @@ from .diffusion_utils import discretized_gaussian_log_likelihood, normal_kl -def mean_flat(tensor, mask=None): +def mean_flat(tensor): """ Take the mean over all non-batch dimensions. """ - if mask is not None: - tensor = tensor * mask return tensor.mean(dim=list(range(1, len(tensor.shape)))) @@ -189,7 +193,7 @@ def __init__(self, *, betas, model_mean_type, model_var_type, loss_type): def q_mean_variance(self, x_start, t): """ Get the distribution q(x_t | x_0). - :param x_start: the [N x T x C x ...] tensor of noiseless inputs. + :param x_start: the [N x C x ...] tensor of noiseless inputs. :param t: the number of diffusion steps (minus 1). Here, 0 means one step. :return: A tuple (mean, variance, log_variance), all of x_start's shape. """ @@ -241,7 +245,7 @@ def p_mean_variance(self, model, x, t, clip_denoised=True, denoised_fn=None, mod the initial x, x_0. :param model: the model, which takes a signal and a batch of timesteps as input. - :param x: the [N x T x C x ...] tensor at time t. + :param x: the [N x C x ...] tensor at time t. :param t: a 1-D Tensor of timesteps. :param clip_denoised: if True, clip the denoised signal into [-1, 1]. :param denoised_fn: if not None, a function which applies to the @@ -258,7 +262,7 @@ def p_mean_variance(self, model, x, t, clip_denoised=True, denoised_fn=None, mod if model_kwargs is None: model_kwargs = {} - B, S, C = x.shape[:3] + B, C = x.shape[:2] assert t.shape == (B,) model_output = model(x, t, **model_kwargs) if isinstance(model_output, tuple): @@ -267,8 +271,8 @@ def p_mean_variance(self, model, x, t, clip_denoised=True, denoised_fn=None, mod extra = None if self.model_var_type in [ModelVarType.LEARNED, ModelVarType.LEARNED_RANGE]: - assert model_output.shape == (B, S, C * 2, *x.shape[3:]) - model_output, model_var_values = th.split(model_output, C, dim=2) + assert model_output.shape == (B, C * 2, *x.shape[2:]) + model_output, model_var_values = th.split(model_output, C, dim=1) min_log = _extract_into_tensor(self.posterior_log_variance_clipped, t, x.shape) max_log = _extract_into_tensor(np.log(self.betas), t, x.shape) # The model_var_values is [-1, 1] for [min_var, max_var]. @@ -411,7 +415,7 @@ def p_sample_loop( """ Generate samples from the model. :param model: the model module. - :param shape: the shape of the samples, (N, T, C, H, W). + :param shape: the shape of the samples, (N, C, H, W). :param noise: if specified, the noise from the encoder to sample. Should be of the same shape as `shape`. :param clip_denoised: if True, clip x_start predictions to [-1, 1]. @@ -649,7 +653,7 @@ def ddim_sample_loop_progressive( yield out img = out["sample"] - def _vb_terms_bpd(self, model, x_start, x_t, t, clip_denoised=True, model_kwargs=None, mask=None): + def _vb_terms_bpd(self, model, x_start, x_t, t, clip_denoised=True, model_kwargs=None): """ Get a term for the variational lower-bound. The resulting units are bits (rather than nats, as one might expect). @@ -661,30 +665,24 @@ def _vb_terms_bpd(self, model, x_start, x_t, t, clip_denoised=True, model_kwargs true_mean, _, true_log_variance_clipped = self.q_posterior_mean_variance(x_start=x_start, x_t=x_t, t=t) out = self.p_mean_variance(model, x_t, t, clip_denoised=clip_denoised, model_kwargs=model_kwargs) kl = normal_kl(true_mean, true_log_variance_clipped, out["mean"], out["log_variance"]) - kl = mean_flat(kl, mask=mask) / np.log(2.0) + kl = mean_flat(kl) / np.log(2.0) decoder_nll = -discretized_gaussian_log_likelihood( x_start, means=out["mean"], log_scales=0.5 * out["log_variance"] ) assert decoder_nll.shape == x_start.shape - decoder_nll = mean_flat(decoder_nll, mask=mask) / np.log(2.0) + decoder_nll = mean_flat(decoder_nll) / np.log(2.0) # At the first timestep return the decoder NLL, # otherwise return KL(q(x_{t-1}|x_t,x_0) || p(x_{t-1}|x_t)) output = th.where((t == 0), decoder_nll, kl) return {"output": output, "pred_xstart": out["pred_xstart"]} - def _expand_mask(self, mask, ndim: int): - assert mask.ndim == 2 - # [B, S] -> [B, S, ...] - mask = mask.view(*mask.shape, *([1] * (ndim - mask.ndim))) - return mask - - def training_losses(self, model, x_start, t, model_kwargs=None, noise=None, mask=None): + def training_losses(self, model, x_start, t, model_kwargs=None, noise=None): """ Compute training losses for a single timestep. :param model: the model to evaluate loss on. - :param x_start: the [N x T x C x ...] tensor of inputs. + :param x_start: the [N x C x ...] tensor of inputs. :param t: a batch of timestep indices. :param model_kwargs: if not None, a dict of extra keyword arguments to pass to the model. This can be used for conditioning. @@ -700,9 +698,6 @@ def training_losses(self, model, x_start, t, model_kwargs=None, noise=None, mask terms = {} - if mask is not None: - mask = self._expand_mask(mask, x_start.ndim) - if self.loss_type == LossType.KL or self.loss_type == LossType.RESCALED_KL: terms["loss"] = self._vb_terms_bpd( model=model, @@ -711,7 +706,6 @@ def training_losses(self, model, x_start, t, model_kwargs=None, noise=None, mask t=t, clip_denoised=False, model_kwargs=model_kwargs, - mask=mask, )["output"] if self.loss_type == LossType.RESCALED_KL: terms["loss"] *= self.num_timesteps @@ -722,19 +716,18 @@ def training_losses(self, model, x_start, t, model_kwargs=None, noise=None, mask ModelVarType.LEARNED, ModelVarType.LEARNED_RANGE, ]: - B, S, C = x_t.shape[:3] - assert model_output.shape == (B, S, C * 2, *x_t.shape[3:]) - model_output, model_var_values = th.split(model_output, C, dim=2) + B, C = x_t.shape[:2] + assert model_output.shape == (B, C * 2, *x_t.shape[2:]) + model_output, model_var_values = th.split(model_output, C, dim=1) # Learn the variance using the variational bound, but don't let # it affect our mean prediction. - frozen_out = th.cat([model_output.detach(), model_var_values], dim=2) + frozen_out = th.cat([model_output.detach(), model_var_values], dim=1) terms["vb"] = self._vb_terms_bpd( model=lambda *args, r=frozen_out: r, x_start=x_start, x_t=x_t, t=t, clip_denoised=False, - mask=mask, )["output"] if self.loss_type == LossType.RESCALED_MSE: # Divide by 1000 for equivalence with initial implementation. @@ -747,7 +740,7 @@ def training_losses(self, model, x_start, t, model_kwargs=None, noise=None, mask ModelMeanType.EPSILON: noise, }[self.model_mean_type] assert model_output.shape == target.shape == x_start.shape - terms["mse"] = mean_flat((target - model_output) ** 2, mask=mask) + terms["mse"] = mean_flat((target - model_output) ** 2) if "vb" in terms: terms["loss"] = terms["mse"] + terms["vb"] else: @@ -762,7 +755,7 @@ def _prior_bpd(self, x_start): Get the prior KL term for the variational lower-bound, measured in bits-per-dim. This term can't be optimized, as it only depends on the encoder. - :param x_start: the [N x T x C x ...] tensor of inputs. + :param x_start: the [N x C x ...] tensor of inputs. :return: a batch of [N] KL values (in bits), one per batch element. """ batch_size = x_start.shape[0] @@ -776,7 +769,7 @@ def calc_bpd_loop(self, model, x_start, clip_denoised=True, model_kwargs=None): Compute the entire variational lower-bound, measured in bits-per-dim, as well as other related quantities. :param model: the model to evaluate loss on. - :param x_start: the [N x T x C x ...] tensor of inputs. + :param x_start: the [N x C x ...] tensor of inputs. :param clip_denoised: if True, clip denoised samples. :param model_kwargs: if not None, a dict of extra keyword arguments to pass to the model. This can be used for conditioning. diff --git a/open_sora/diffusion/respace.py b/opensora/schedulers/iddpm/respace.py similarity index 89% rename from open_sora/diffusion/respace.py rename to opensora/schedulers/iddpm/respace.py index e5754aa7..d5ea16ce 100644 --- a/open_sora/diffusion/respace.py +++ b/opensora/schedulers/iddpm/respace.py @@ -1,7 +1,15 @@ -# Modified from OpenAI's diffusion repos -# GLIDE: https://github.com/openai/glide-text2im/blob/main/glide_text2im/gaussian_diffusion.py -# ADM: https://github.com/openai/guided-diffusion/blob/main/guided_diffusion -# IDDPM: https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py +# Adapted from DiT + +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. +# -------------------------------------------------------- +# References: +# DiT: https://github.com/facebookresearch/DiT/tree/main +# GLIDE: https://github.com/openai/glide-text2im/blob/main/glide_text2im/gaussian_diffusion.py +# ADM: https://github.com/openai/guided-diffusion/blob/main/guided_diffusion +# IDDPM: https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py +# -------------------------------------------------------- + import numpy as np import torch as th diff --git a/open_sora/diffusion/timestep_sampler.py b/opensora/schedulers/iddpm/timestep_sampler.py similarity index 90% rename from open_sora/diffusion/timestep_sampler.py rename to opensora/schedulers/iddpm/timestep_sampler.py index fdaa45ac..52b6717d 100644 --- a/open_sora/diffusion/timestep_sampler.py +++ b/opensora/schedulers/iddpm/timestep_sampler.py @@ -1,7 +1,14 @@ -# Modified from OpenAI's diffusion repos -# GLIDE: https://github.com/openai/glide-text2im/blob/main/glide_text2im/gaussian_diffusion.py -# ADM: https://github.com/openai/guided-diffusion/blob/main/guided_diffusion -# IDDPM: https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py +# Adapted from DiT + +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. +# -------------------------------------------------------- +# References: +# DiT: https://github.com/facebookresearch/DiT/tree/main +# GLIDE: https://github.com/openai/glide-text2im/blob/main/glide_text2im/gaussian_diffusion.py +# ADM: https://github.com/openai/guided-diffusion/blob/main/guided_diffusion +# IDDPM: https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py +# -------------------------------------------------------- from abc import ABC, abstractmethod diff --git a/opensora/utils/__init__.py b/opensora/utils/__init__.py new file mode 100644 index 00000000..e69de29b diff --git a/opensora/utils/ckpt_utils.py b/opensora/utils/ckpt_utils.py new file mode 100644 index 00000000..27adfba1 --- /dev/null +++ b/opensora/utils/ckpt_utils.py @@ -0,0 +1,216 @@ +import functools +import json +import logging +import operator +import os +from typing import Tuple + +import colossalai +import torch +import torch.distributed as dist +import torch.nn as nn +from colossalai.booster import Booster +from colossalai.checkpoint_io import GeneralCheckpointIO +from colossalai.cluster import DistCoordinator +from torch.optim import Optimizer +from torch.optim.lr_scheduler import _LRScheduler +from torchvision.datasets.utils import download_url + +pretrained_models = { + "DiT-XL-2-512x512.pt": "https://dl.fbaipublicfiles.com/DiT/models/DiT-XL-2-512x512.pt", + "DiT-XL-2-256x256.pt": "https://dl.fbaipublicfiles.com/DiT/models/DiT-XL-2-256x256.pt", + "Latte-XL-2-256x256-ucf101.pt": "https://huggingface.co/maxin-cn/Latte/resolve/main/ucf101.pt", + "PixArt-XL-2-256x256.pth": "https://huggingface.co/PixArt-alpha/PixArt-alpha/resolve/main/PixArt-XL-2-256x256.pth", + "PixArt-XL-2-SAM-256x256.pth": "https://huggingface.co/PixArt-alpha/PixArt-alpha/resolve/main/PixArt-XL-2-SAM-256x256.pth", + "PixArt-XL-2-512x512.pth": "https://huggingface.co/PixArt-alpha/PixArt-alpha/resolve/main/PixArt-XL-2-512x512.pth", + "PixArt-XL-2-1024-MS.pth": "https://huggingface.co/PixArt-alpha/PixArt-alpha/resolve/main/PixArt-XL-2-1024-MS.pth", +} + + +def reparameter(ckpt, name=None): + if "DiT" in name: + ckpt["x_embedder.proj.weight"] = ckpt["x_embedder.proj.weight"].unsqueeze(2) + del ckpt["pos_embed"] + elif "Latte" in name: + ckpt = ckpt["ema"] + ckpt["x_embedder.proj.weight"] = ckpt["x_embedder.proj.weight"].unsqueeze(2) + del ckpt["pos_embed"] + del ckpt["temp_embed"] + elif "PixArt" in name: + ckpt = ckpt["state_dict"] + ckpt["x_embedder.proj.weight"] = ckpt["x_embedder.proj.weight"].unsqueeze(2) + del ckpt["pos_embed"] + return ckpt + + +def find_model(model_name): + """ + Finds a pre-trained DiT model, downloading it if necessary. Alternatively, loads a model from a local path. + """ + if model_name in pretrained_models: # Find/download our pre-trained DiT checkpoints + model = download_model(model_name) + model = reparameter(model, model_name) + return model + else: # Load a custom DiT checkpoint: + assert os.path.isfile(model_name), f"Could not find DiT checkpoint at {model_name}" + checkpoint = torch.load(model_name, map_location=lambda storage, loc: storage) + if "pos_embed_temporal" in checkpoint: + del checkpoint["pos_embed_temporal"] + if "pos_embed" in checkpoint: + del checkpoint["pos_embed"] + if "ema" in checkpoint: # supports checkpoints from train.py + checkpoint = checkpoint["ema"] + return checkpoint + + +def download_model(model_name): + """ + Downloads a pre-trained DiT model from the web. + """ + assert model_name in pretrained_models + local_path = f"pretrained_models/{model_name}" + if not os.path.isfile(local_path): + os.makedirs("pretrained_models", exist_ok=True) + web_path = pretrained_models[model_name] + download_url(web_path, "pretrained_models", model_name) + model = torch.load(local_path, map_location=lambda storage, loc: storage) + return model + + +def load_from_sharded_state_dict(model, ckpt_path): + ckpt_io = GeneralCheckpointIO() + ckpt_io.load_model(model, os.path.join(ckpt_path, "model")) + +def model_sharding(model: torch.nn.Module): + global_rank = dist.get_rank() + world_size = dist.get_world_size() + for _, param in model.named_parameters(): + padding_size = (world_size - param.numel() % world_size) % world_size + if padding_size > 0: + padding_param = torch.nn.functional.pad(param.data.view(-1), [0, padding_size]) + else: + padding_param = param.data.view(-1) + splited_params = padding_param.split(padding_param.numel() // world_size) + splited_params = splited_params[global_rank] + param.data = splited_params + + +def load_json(file_path: str): + with open(file_path, "r") as f: + return json.load(f) + + +def save_json(data, file_path: str): + with open(file_path, "w") as f: + json.dump(data, f, indent=4) + + +def remove_padding(tensor: torch.Tensor, original_shape: Tuple) -> torch.Tensor: + return tensor[: functools.reduce(operator.mul, original_shape)] + + +def model_gathering(model: torch.nn.Module, model_shape_dict: dict): + global_rank = dist.get_rank() + global_size = dist.get_world_size() + for name, param in model.named_parameters(): + all_params = [torch.empty_like(param.data) for _ in range(global_size)] + dist.all_gather(all_params, param.data, group=dist.group.WORLD) + if int(global_rank) == 0: + all_params = torch.cat(all_params) + param.data = remove_padding(all_params, model_shape_dict[name]).view(model_shape_dict[name]) + dist.barrier() + + +def record_model_param_shape(model: torch.nn.Module) -> dict: + param_shape = {} + for name, param in model.named_parameters(): + param_shape[name] = param.shape + return param_shape + + +def save( + booster: Booster, + model: nn.Module, + ema: nn.Module, + optimizer: Optimizer, + lr_scheduler: _LRScheduler, + epoch: int, + step: int, + global_step: int, + batch_size: int, + coordinator: DistCoordinator, + save_dir: str, + shape_dict: dict, +): + save_dir = os.path.join(save_dir, f"epoch{epoch}-global_step{global_step}") + os.makedirs(os.path.join(save_dir, "model"), exist_ok=True) + + booster.save_model(model, os.path.join(save_dir, "model"), shard=True) + # ema is not boosted, so we don't need to use booster.save_model + model_gathering(ema, shape_dict) + global_rank = dist.get_rank() + if int(global_rank) == 0: + torch.save(ema.state_dict(), os.path.join(save_dir, "ema.pt")) + model_sharding(ema) + + booster.save_optimizer(optimizer, os.path.join(save_dir, "optimizer"), shard=True, size_per_shard=4096) + if lr_scheduler is not None: + booster.save_lr_scheduler(lr_scheduler, os.path.join(save_dir, "lr_scheduler")) + running_states = { + "epoch": epoch, + "step": step, + "global_step": global_step, + "sample_start_index": step * batch_size, + } + if coordinator.is_master(): + save_json(running_states, os.path.join(save_dir, "running_states.json")) + dist.barrier() + + +def load( + booster: Booster, model: nn.Module, ema: nn.Module, optimizer: Optimizer, lr_scheduler: _LRScheduler, load_dir: str +) -> Tuple[int, int, int]: + booster.load_model(model, os.path.join(load_dir, "model")) + # ema is not boosted, so we don't use booster.load_model + # ema.load_state_dict(torch.load(os.path.join(load_dir, "ema.pt"))) + ema.load_state_dict(torch.load(os.path.join(load_dir, "ema.pt"), map_location=torch.device("cpu"))) + booster.load_optimizer(optimizer, os.path.join(load_dir, "optimizer")) + if lr_scheduler is not None: + booster.load_lr_scheduler(lr_scheduler, os.path.join(load_dir, "lr_scheduler")) + running_states = load_json(os.path.join(load_dir, "running_states.json")) + dist.barrier() + return running_states["epoch"], running_states["step"], running_states["sample_start_index"] + + +def create_logger(logging_dir): + """ + Create a logger that writes to a log file and stdout. + """ + if dist.get_rank() == 0: # real logger + logging.basicConfig( + level=logging.INFO, + format="[\033[34m%(asctime)s\033[0m] %(message)s", + datefmt="%Y-%m-%d %H:%M:%S", + handlers=[logging.StreamHandler(), logging.FileHandler(f"{logging_dir}/log.txt")], + ) + logger = logging.getLogger(__name__) + else: # dummy logger (does nothing) + logger = logging.getLogger(__name__) + logger.addHandler(logging.NullHandler()) + return logger + + +def load_checkpoint(model, ckpt_path, save_as_pt=True): + if ckpt_path.endswith(".pt") or ckpt_path.endswith(".pth"): + state_dict = find_model(ckpt_path) + missing_keys, unexpected_keys = model.load_state_dict(state_dict, strict=False) + print(f"Missing keys: {missing_keys}") + print(f"Unexpected keys: {unexpected_keys}") + elif os.path.isdir(ckpt_path): + load_from_sharded_state_dict(model, ckpt_path) + if save_as_pt: + save_path = os.path.join(ckpt_path, "model_ckpt.pt") + torch.save(model.state_dict(), save_path) + print(f"Model checkpoint saved to {save_path}") + else: + raise ValueError(f"Invalid checkpoint path: {ckpt_path}") diff --git a/opensora/utils/config_utils.py b/opensora/utils/config_utils.py new file mode 100644 index 00000000..5ef81506 --- /dev/null +++ b/opensora/utils/config_utils.py @@ -0,0 +1,97 @@ +import argparse +import json +import os +from glob import glob + +from mmengine.config import Config +from torch.utils.tensorboard import SummaryWriter + + +def parse_args(training=False): + parser = argparse.ArgumentParser() + + # model config + parser.add_argument("config", help="model config file path") + + parser.add_argument("--seed", default=42, type=int, help="generation seed") + parser.add_argument("--ckpt-path", type=str, help="path to model ckpt; will overwrite cfg.ckpt_path if specified") + parser.add_argument("--batch-size", default=None, type=int, help="batch size") + + # ====================================================== + # Inference + # ====================================================== + + if not training: + # prompt + parser.add_argument("--prompt-path", default=None, type=str, help="path to prompt txt file") + parser.add_argument("--save-dir", default=None, type=str, help="path to save generated samples") + + # hyperparameters + parser.add_argument("--num-sampling-steps", default=None, type=int, help="sampling steps") + parser.add_argument("--cfg-scale", default=None, type=float, help="balance between cond & uncond") + else: + parser.add_argument("--wandb", default=None, type=bool, help="enable wandb") + parser.add_argument("--load", default=None, type=str, help="path to continue training") + parser.add_argument("--data-path", default=None, type=str, help="path to data csv") + + return parser.parse_args() + + +def merge_args(cfg, args, training=False): + if args.ckpt_path is not None: + cfg.model["from_pretrained"] = args.ckpt_path + args.ckpt_path = None + + if not training: + if args.cfg_scale is not None: + cfg.scheduler["cfg_scale"] = args.cfg_scale + args.cfg_scale = None + + if "multi_resolution" not in cfg: + cfg["multi_resolution"] = False + for k, v in vars(args).items(): + if k in cfg and v is not None: + cfg[k] = v + + return cfg + + +def parse_configs(training=False): + args = parse_args(training) + cfg = Config.fromfile(args.config) + cfg = merge_args(cfg, args, training) + return cfg + + +def create_experiment_workspace(cfg): + """ + This function creates a folder for experiment tracking. + + Args: + args: The parsed arguments. + + Returns: + exp_dir: The path to the experiment folder. + """ + # Make outputs folder (holds all experiment subfolders) + os.makedirs(cfg.outputs, exist_ok=True) + experiment_index = len(glob(f"{cfg.outputs}/*")) + + # Create an experiment folder + model_name = cfg.model["type"].replace("/", "-") + exp_name = f"{experiment_index:03d}-F{cfg.num_frames}S{cfg.frame_interval}-{model_name}" + exp_dir = f"{cfg.outputs}/{exp_name}" + os.makedirs(exp_dir, exist_ok=True) + return exp_name, exp_dir + + +def save_training_config(cfg, experiment_dir): + with open(f"{experiment_dir}/config.txt", "w") as f: + json.dump(cfg, f, indent=4) + + +def create_tensorboard_writer(exp_dir): + tensorboard_dir = f"{exp_dir}/tensorboard" + os.makedirs(tensorboard_dir, exist_ok=True) + writer = SummaryWriter(tensorboard_dir) + return writer diff --git a/opensora/utils/misc.py b/opensora/utils/misc.py new file mode 100644 index 00000000..d1625261 --- /dev/null +++ b/opensora/utils/misc.py @@ -0,0 +1,286 @@ +import collections +import importlib +import logging +import os +import time +from collections import OrderedDict +from collections.abc import Sequence +from itertools import repeat + +import numpy as np +import torch +import torch.distributed as dist + + +def print_rank(var_name, var_value, rank=0): + if dist.get_rank() == rank: + print(f"[Rank {rank}] {var_name}: {var_value}") + + +def print_0(*args, **kwargs): + if dist.get_rank() == 0: + print(*args, **kwargs) + + +def requires_grad(model: torch.nn.Module, flag: bool = True) -> None: + """ + Set requires_grad flag for all parameters in a model. + """ + for p in model.parameters(): + p.requires_grad = flag + + +def format_numel_str(numel: int) -> str: + B = 1024**3 + M = 1024**2 + K = 1024 + if numel >= B: + return f"{numel / B:.2f} B" + elif numel >= M: + return f"{numel / M:.2f} M" + elif numel >= K: + return f"{numel / K:.2f} K" + else: + return f"{numel}" + + +def all_reduce_mean(tensor: torch.Tensor) -> torch.Tensor: + dist.all_reduce(tensor=tensor, op=dist.ReduceOp.SUM) + tensor.div_(dist.get_world_size()) + return tensor + + +def get_model_numel(model: torch.nn.Module) -> (int, int): + num_params = 0 + num_params_trainable = 0 + for p in model.parameters(): + num_params += p.numel() + if p.requires_grad: + num_params_trainable += p.numel() + return num_params, num_params_trainable + + +def try_import(name): + """Try to import a module. + + Args: + name (str): Specifies what module to import in absolute or relative + terms (e.g. either pkg.mod or ..mod). + Returns: + ModuleType or None: If importing successfully, returns the imported + module, otherwise returns None. + """ + try: + return importlib.import_module(name) + except ImportError: + return None + + +def transpose(x): + """ + transpose a list of list + Args: + x (list[list]): + """ + ret = list(map(list, zip(*x))) + return ret + + +def get_timestamp(): + timestamp = time.strftime("%Y%m%d-%H%M%S", time.localtime(time.time())) + return timestamp + + +def format_time(seconds): + days = int(seconds / 3600 / 24) + seconds = seconds - days * 3600 * 24 + hours = int(seconds / 3600) + seconds = seconds - hours * 3600 + minutes = int(seconds / 60) + seconds = seconds - minutes * 60 + secondsf = int(seconds) + seconds = seconds - secondsf + millis = int(seconds * 1000) + + f = "" + i = 1 + if days > 0: + f += str(days) + "D" + i += 1 + if hours > 0 and i <= 2: + f += str(hours) + "h" + i += 1 + if minutes > 0 and i <= 2: + f += str(minutes) + "m" + i += 1 + if secondsf > 0 and i <= 2: + f += str(secondsf) + "s" + i += 1 + if millis > 0 and i <= 2: + f += str(millis) + "ms" + i += 1 + if f == "": + f = "0ms" + return f + + +def to_tensor(data): + """Convert objects of various python types to :obj:`torch.Tensor`. + + Supported types are: :class:`numpy.ndarray`, :class:`torch.Tensor`, + :class:`Sequence`, :class:`int` and :class:`float`. + + Args: + data (torch.Tensor | numpy.ndarray | Sequence | int | float): Data to + be converted. + """ + + if isinstance(data, torch.Tensor): + return data + elif isinstance(data, np.ndarray): + return torch.from_numpy(data) + elif isinstance(data, Sequence) and not isinstance(data, str): + return torch.tensor(data) + elif isinstance(data, int): + return torch.LongTensor([data]) + elif isinstance(data, float): + return torch.FloatTensor([data]) + else: + raise TypeError(f"type {type(data)} cannot be converted to tensor.") + + +def to_ndarray(data): + if isinstance(data, torch.Tensor): + return data.numpy() + elif isinstance(data, np.ndarray): + return data + elif isinstance(data, Sequence): + return np.array(data) + elif isinstance(data, int): + return np.ndarray([data], dtype=int) + elif isinstance(data, float): + return np.array([data], dtype=float) + else: + raise TypeError(f"type {type(data)} cannot be converted to ndarray.") + + +def to_torch_dtype(dtype): + if isinstance(dtype, torch.dtype): + return dtype + elif isinstance(dtype, str): + dtype_mapping = { + "float64": torch.float64, + "float32": torch.float32, + "float16": torch.float16, + "fp32": torch.float32, + "fp16": torch.float16, + "half": torch.float16, + "bf16": torch.bfloat16, + } + if dtype not in dtype_mapping: + raise ValueError + dtype = dtype_mapping[dtype] + return dtype + else: + raise ValueError + + +def count_params(model): + return sum(p.numel() for p in model.parameters() if p.requires_grad) + + +def _ntuple(n): + def parse(x): + if isinstance(x, collections.abc.Iterable) and not isinstance(x, str): + return x + return tuple(repeat(x, n)) + + return parse + + +to_1tuple = _ntuple(1) +to_2tuple = _ntuple(2) +to_3tuple = _ntuple(3) +to_4tuple = _ntuple(4) +to_ntuple = _ntuple + + +def convert_SyncBN_to_BN2d(model_cfg): + for k in model_cfg: + v = model_cfg[k] + if k == "norm_cfg" and v["type"] == "SyncBN": + v["type"] = "BN2d" + elif isinstance(v, dict): + convert_SyncBN_to_BN2d(v) + + +def get_topk(x, dim=4, k=5): + x = to_tensor(x) + inds = x[..., dim].topk(k)[1] + return x[inds] + + +def param_sigmoid(x, alpha): + ret = 1 / (1 + (-alpha * x).exp()) + return ret + + +def inverse_param_sigmoid(x, alpha, eps=1e-5): + x = x.clamp(min=0, max=1) + x1 = x.clamp(min=eps) + x2 = (1 - x).clamp(min=eps) + return torch.log(x1 / x2) / alpha + + +def inverse_sigmoid(x, eps=1e-5): + """Inverse function of sigmoid. + + Args: + x (Tensor): The tensor to do the + inverse. + eps (float): EPS avoid numerical + overflow. Defaults 1e-5. + Returns: + Tensor: The x has passed the inverse + function of sigmoid, has same + shape with input. + """ + x = x.clamp(min=0, max=1) + x1 = x.clamp(min=eps) + x2 = (1 - x).clamp(min=eps) + return torch.log(x1 / x2) + + +def count_columns(df, columns): + cnt_dict = OrderedDict() + num_samples = len(df) + + for col in columns: + d_i = df[col].value_counts().to_dict() + for k in d_i: + d_i[k] = (d_i[k], d_i[k] / num_samples) + cnt_dict[col] = d_i + + return cnt_dict + + +def build_logger(work_dir, cfgname): + log_file = cfgname + ".log" + log_path = os.path.join(work_dir, log_file) + + logger = logging.getLogger(cfgname) + logger.setLevel(logging.INFO) + # formatter = logging.Formatter('%(asctime)s %(levelname)s %(message)s') + formatter = logging.Formatter("%(asctime)s: %(message)s", datefmt="%Y-%m-%d %H:%M:%S") + + handler1 = logging.FileHandler(log_path) + handler1.setFormatter(formatter) + + handler2 = logging.StreamHandler() + handler2.setFormatter(formatter) + + logger.addHandler(handler1) + logger.addHandler(handler2) + logger.propagate = False + + return logger diff --git a/opensora/utils/train_utils.py b/opensora/utils/train_utils.py new file mode 100644 index 00000000..f8460439 --- /dev/null +++ b/opensora/utils/train_utils.py @@ -0,0 +1,31 @@ +from collections import OrderedDict + +import torch + + +@torch.no_grad() +def update_ema( + ema_model: torch.nn.Module, model: torch.nn.Module, optimizer=None, decay: float = 0.9999, sharded: bool = True +) -> None: + """ + Step the EMA model towards the current model. + """ + ema_params = OrderedDict(ema_model.named_parameters()) + model_params = OrderedDict(model.named_parameters()) + + for name, param in model_params.items(): + if name == "pos_embed": + continue + if param.requires_grad == False: + continue + if not sharded: + param_data = param.data + ema_params[name].mul_(decay).add_(param_data, alpha=1 - decay) + else: + if param.data.dtype != torch.float32: + param_id = id(param) + master_param = optimizer._param_store.working_to_master_param[param_id] + param_data = master_param.data + else: + param_data = param.data + ema_params[name].mul_(decay).add_(param_data, alpha=1 - decay) diff --git a/requirements.txt b/requirements.txt index 3a8077d5..0724a86b 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,9 +1,13 @@ -torch -torchvision -datasets -transformers -av +colossalai +accelerate +diffusers +ftfy +gdown +mmengine +pre-commit +pyav tensorboard -git+https://github.com/hpcaitech/ColossalAI.git@main#egg=colossalai timm -diffusers==0.24.0 \ No newline at end of file +tqdm +transformers +wandb diff --git a/sample.py b/sample.py deleted file mode 100644 index 67aec4bf..00000000 --- a/sample.py +++ /dev/null @@ -1,150 +0,0 @@ -# Copyright (c) Meta Platforms, Inc. and affiliates. -# All rights reserved. - -# This source code is licensed under the license found in the -# LICENSE file in the root directory of this source tree. - -""" -Sample new images from a pre-trained DiT. -""" -import torch - -torch.backends.cuda.matmul.allow_tf32 = True -torch.backends.cudnn.allow_tf32 = True -import argparse - -from colossalai.utils import get_current_device -from torchvision.io import write_video -from transformers import AutoTokenizer, CLIPTextModel - -from open_sora.diffusion import create_diffusion -from open_sora.modeling import DiT_models -from open_sora.modeling.dit import SUPPORTED_MODEL_ARCH -from open_sora.utils.data import col2video, create_video_compressor, unnormalize_video - - -def main(args): - # Setup PyTorch: - torch.manual_seed(args.seed) - torch.set_grad_enabled(False) - device = get_current_device() - - video_compressor = create_video_compressor(args.compressor) - model_kwargs = { - "in_channels": video_compressor.out_channels, - "model_arch": args.model_arch, - } - - text_model = CLIPTextModel.from_pretrained(args.text_model).to(device).eval() - tokenizer = AutoTokenizer.from_pretrained(args.text_model) - - model = DiT_models[args.model](**model_kwargs).to(device).eval() - patch_size = model.patch_size - model.load_state_dict(torch.load(args.ckpt)) - diffusion = create_diffusion(str(args.num_sampling_steps)) - - # Create sampling noise: - text_inputs = tokenizer(args.text, return_tensors="pt") - text_inputs = {k: v.to(device) for k, v in text_inputs.items()} - text_latent_states = text_model(**text_inputs) - if args.model_arch == "adaln": - text_latent_states = text_latent_states.pooler_output - else: - text_latent_states = text_latent_states.last_hidden_state - - num_frames = args.fps * args.sec - z = torch.randn( - 1, - (args.height // patch_size // video_compressor.h_w_factor) - * (args.width // patch_size // video_compressor.h_w_factor) - * (num_frames // video_compressor.t_factor), - video_compressor.out_channels, - patch_size, - patch_size, - device=device, - ) - - # Setup classifier-free guidance: - model_kwargs = {} - if not args.disable_cfg: - z = torch.cat([z, z], 0) - model_kwargs["text_latent_states"] = torch.cat( - [text_latent_states, torch.zeros_like(text_latent_states)], 0 - ) - model_kwargs["cfg_scale"] = args.cfg_scale - else: - model_kwargs["text_latent_states"] = text_latent_states - context_len = ( - text_latent_states.shape[1] if args.model_arch == "cross-attn" else z.shape[1] - ) - model_kwargs["attention_mask"] = torch.ones( - z.shape[0], - 1, - z.shape[1], - context_len, - device=device, - dtype=torch.int, - ) - - # Sample images: - samples = diffusion.p_sample_loop( - model if args.disable_cfg else model.forward_with_cfg, - z.shape, - z, - clip_denoised=False, - model_kwargs=model_kwargs, - progress=True, - device=device, - ) - if not args.disable_cfg: - samples, _ = samples.chunk(2, dim=0) # Remove null class samples - samples = col2video( - samples.squeeze(), - ( - num_frames // video_compressor.t_factor, - video_compressor.out_channels, - args.height // video_compressor.h_w_factor, - args.width // video_compressor.h_w_factor, - ), - ) - samples = video_compressor.decode(samples) - samples = unnormalize_video(samples).to(torch.uint8) - - write_video("sample.mp4", samples.cpu(), args.fps) - - -if __name__ == "__main__": - parser = argparse.ArgumentParser() - parser.add_argument( - "-m", "--model", type=str, choices=list(DiT_models.keys()), default="DiT-S/8" - ) - parser.add_argument( - "-x", "--model_arch", choices=SUPPORTED_MODEL_ARCH, default="cross-attn" - ) - parser.add_argument( - "--text", - type=str, - default="a cartoon animals runs through an ice cave in a video game", - ) - parser.add_argument("--cfg-scale", type=float, default=4.0) - parser.add_argument("--num-sampling-steps", type=int, default=250) - parser.add_argument("--seed", type=int, default=0) - parser.add_argument( - "--ckpt", - type=str, - required=True, - help="Optional path to a DiT checkpoint (default: auto-download a pre-trained DiT-XL/2 model).", - ) - parser.add_argument( - "-c", "--compressor", choices=["raw", "vqvae", "vae"], default="raw" - ) - parser.add_argument( - "--text_model", type=str, default="openai/clip-vit-base-patch32" - ) - parser.add_argument("--width", type=int, default=480) - parser.add_argument("--height", type=int, default=320) - parser.add_argument("--fps", type=int, default=15) - parser.add_argument("--sec", type=int, default=8) - parser.add_argument("--disable-cfg", action="store_true", default=False) - args = parser.parse_args() - main(args) diff --git a/scripts/data/collate_msr_vtt_dataset.py b/scripts/data/collate_msr_vtt_dataset.py deleted file mode 100644 index 04f63cbd..00000000 --- a/scripts/data/collate_msr_vtt_dataset.py +++ /dev/null @@ -1,182 +0,0 @@ -import argparse -import json -import multiprocessing -import os -import shutil -import warnings -from typing import Dict, Tuple - -from tqdm import tqdm - -DEFAULT_TYPES = ["train", "val", "test"] - - -def parse_args(): - parser = argparse.ArgumentParser() - parser.add_argument("-d", "--data-path", type=str, help="The path to the MSR-VTT dataset") - parser.add_argument("-o", "--output-path", type=str, help="The output to the collated MSR-VTT dataset") - return parser.parse_args() - - -def get_annotations(root_path: str): - """ - Get the annotation data from the MSR-VTT dataset. The annotations are in the format of: - - { - "annotations": [ - { - "image_id": "video1", - "caption": "some - } - ] - } - - Args: - root_path (str): The root path to the MSR-VTT dataset - """ - annotation_json_file = os.path.join(root_path, "annotation/MSR_VTT.json") - with open(annotation_json_file, "r") as f: - data = json.load(f) - return data - - -def get_video_list(root_path: str, dataset_type: str): - """ - Get the list of videos in the dataset split. - - Args: - root_path (str): The root path to the MSR-VTT dataset - dataset_type (str): The dataset split type. It should be one of "train", "val", or "test" - """ - assert dataset_type in DEFAULT_TYPES, f"Expected the dataset type to be in {DEFAULT_TYPES}, but got {dataset_type}" - dataset_file_path = os.path.join(root_path, f"structured-symlinks/{dataset_type}_list_full.txt") - with open(dataset_file_path, "r") as f: - video_list = f.readlines() - video_list = [x.strip() for x in video_list] - return video_list - - -def copy_video(video_id: str, root_path: str, output_path: str, dataset_type: str): - """ - Copy the video from the source path to the destination path. - - Args: - video_id (str): The video id - root_path (str): The root path to the MSR-VTT dataset - output_path (str): The output path to the collated MSR-VTT dataset - dataset_type (str): The dataset split type. It should be one of "train", "val", or "test" - """ - assert dataset_type in DEFAULT_TYPES, f"Expected the dataset type to be in {DEFAULT_TYPES}, but got {dataset_type}" - src_file = os.path.join(root_path, f"videos/all/{video_id}.mp4") - dst_folder = os.path.join(output_path, f"{dataset_type}/videos") - dst_file = os.path.join(dst_folder, f"{video_id}.mp4") - os.makedirs(dst_folder, exist_ok=True) - - # create symlink - assert os.path.isfile(src_file), f"Expected the source file {src_file} to exist" - if not os.path.islink(dst_file): - shutil.copy(src_file, dst_file) - - -def get_annotation_file_path(output_path: str, dataset_type: str): - file_path = os.path.join(output_path, f"{dataset_type}/annotations.json") - return file_path - - -def collate_annotation_files( - annotations: Dict, - root_path: str, - output_path: str, -): - """ - Collate the video and caption data into a single folder. - - Args: - annotations (Dict): The annotations data - root_path (str): The root path to the MSR-VTT dataset - output_path (str): The output path to the collated MSR-VTT dataset - """ - # get all video list - train_video_list = get_video_list(root_path, "train") - val_video_list = get_video_list(root_path, "val") - test_video_list = get_video_list(root_path, "test") - - # iterate over annotations - collated_train_data = [] - collated_val_data = [] - collated_test_data = [] - - print("Collating annotations files") - - for anno in tqdm(annotations["annotations"]): - video_id = anno["image_id"] - caption = anno["caption"] - - obj = {"file": f"{video_id}.mp4", "captions": [caption]} - - if video_id in train_video_list: - collated_train_data.append(obj) - elif video_id in val_video_list: - collated_val_data.append(obj) - elif video_id in test_video_list: - collated_test_data.append(obj) - else: - warnings.warn(f"Video {video_id} not found in any of the dataset splits") - - def _save_caption_files(obj, dataset_type): - dst_file = get_annotation_file_path(output_path, dataset_type) - os.makedirs(os.path.dirname(dst_file), exist_ok=True) - with open(dst_file, "w") as f: - json.dump(obj, f, indent=4) - - _save_caption_files(collated_train_data, "train") - _save_caption_files(collated_val_data, "val") - _save_caption_files(collated_test_data, "test") - - -def copy_file(path_pair: Tuple[str, str]): - src_path, dst_path = path_pair - shutil.copyfile(src_path, dst_path) - - -def copy_videos(root_path: str, output_path: str, num_workers: int = 8): - """ - Batch copy the video files to the output path. - - Args: - root_path (str): The root path to the MSR-VTT dataset - output_path (str): The output path to the collated MSR-VTT dataset - num_workers (int): The number of workers to use for the copy operation - """ - pool = multiprocessing.Pool(num_workers) - - for dataset_type in DEFAULT_TYPES: - print(f"Copying videos for the {dataset_type} dataset") - annotation_file_path = get_annotation_file_path(output_path, dataset_type) - output_video_folder_path = os.path.join(output_path, f"{dataset_type}/videos") - os.makedirs(output_video_folder_path, exist_ok=True) - - with open(annotation_file_path, "r") as f: - annotation_data = json.load(f) - - video_ids = [obj["file"] for obj in annotation_data] - unique_video_ids = list(set(video_ids)) - - path_pairs = [ - (os.path.join(root_path, f"videos/all/{video_id}"), os.path.join(output_video_folder_path, video_id)) - for video_id in unique_video_ids - ] - - for _ in tqdm(pool.imap_unordered(copy_file, path_pairs), total=len(path_pairs)): - pass - - -def main(): - args = parse_args() - annotations = get_annotations(args.data_path) - collate_annotation_files(annotations, args.data_path, args.output_path) - copy_videos(args.data_path, args.output_path) - - -if __name__ == "__main__": - main() diff --git a/scripts/data/download_msr_vtt_dataset.sh b/scripts/data/download_msr_vtt_dataset.sh deleted file mode 100644 index 141a7d62..00000000 --- a/scripts/data/download_msr_vtt_dataset.sh +++ /dev/null @@ -1,11 +0,0 @@ -#!/usr/bin/env bash - -# get root dir -FOLDER_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )" -ROOT_DIR=$FOLDER_DIR/../.. - -# download at root dir -cd $ROOT_DIR -mkdir -p dataset && cd ./dataset -wget https://www.robots.ox.ac.uk/~maxbain/frozen-in-time/data/MSRVTT.zip -unzip MSRVTT.zip diff --git a/scripts/data/preprocess_data.py b/scripts/data/preprocess_data.py deleted file mode 100644 index bc0dfbb7..00000000 --- a/scripts/data/preprocess_data.py +++ /dev/null @@ -1,135 +0,0 @@ -import argparse -import math -import os - -import torch -from datasets import load_dataset -from torchvision.io import read_video -from transformers import AutoTokenizer, CLIPTextModel - -EMPTY_SAMPLE = {"video_file": [], "text_latent_states": []} - - -def process_text(text, tokenizer, text_model, use_pooled_text): - inputs = tokenizer(text, padding=True, return_tensors="pt") - inputs = {k: v.cuda() for k, v in inputs.items()} - outputs = text_model(**inputs) - if use_pooled_text: - return list(outputs.pooler_output.cpu().unbind(0)) - output_states = [] - for i, x in enumerate(outputs.last_hidden_state): - valid_x = x[inputs["attention_mask"][i].bool()] - output_states.append(valid_x.cpu()) - return output_states - - -@torch.no_grad() -def process_item(item, video_dir, tokenizer, text_model, use_pooled_text): - video_path = os.path.join(video_dir, item["file"]) - video = read_video(video_path, pts_unit="sec")[0] - if video.size(0) > 600: - return EMPTY_SAMPLE - text_latent_states = process_text( - item["captions"], tokenizer, text_model, use_pooled_text - ) - torch.cuda.empty_cache() - return { - "video_file": [item["file"]] * len(text_latent_states), - "text_latent_states": text_latent_states, - } - - -def process_batch(batch, video_dir, tokenizer, text_model, use_pooled_text): - item = {"file": batch["file"][0], "captions": batch["captions"][0]} - return process_item(item, video_dir, tokenizer, text_model, use_pooled_text) - - -def process_dataset( - captions_file, - video_dir, - output_dir, - num_spliced_dataset_bins=10, - text_model="openai/clip-vit-base-patch32", - use_pooled_text=False, -): - tokenizer = AutoTokenizer.from_pretrained(text_model) - text_model = CLIPTextModel.from_pretrained(text_model).cuda().eval() - - if not os.path.exists(output_dir): - os.makedirs(output_dir) - - # Prepare to data splitting. - train_splits = [] - split_interval = math.ceil(100 / num_spliced_dataset_bins) - for i in range(0, 100, split_interval): - start = i - end = i + split_interval - if end > 100: - end = 100 - train_splits.append(f"train[{start}%:{end}%]") - - ds = load_dataset( - "json", - data_files=captions_file, - keep_in_memory=False, - split=train_splits, - num_proc=1, - ) - - for i, part_ds in enumerate(ds): - print(f"Processing part {i+1}/{len(ds)}") - part_ds = part_ds.with_format("torch") - part_ds = part_ds.map( - process_batch, - fn_kwargs={ - "video_dir": video_dir, - "tokenizer": tokenizer, - "text_model": text_model, - "use_pooled_text": use_pooled_text, - }, - batched=True, - batch_size=1, - keep_in_memory=False, - remove_columns=part_ds.column_names, - ) - output_path = os.path.join(output_dir, f"part-{i:05d}") - part_ds.save_to_disk(output_path) - - -if __name__ == "__main__": - parser = argparse.ArgumentParser(description="Preprocess data") - parser.add_argument( - "-c", - "--captions-file", - type=str, - help="Path to the captions file. It should be a JSON file or a JSONL file", - ) - parser.add_argument( - "-v", "--video-dir", type=str, help="Path to the video directory" - ) - parser.add_argument( - "-o", "--output_dir", type=str, help="Path to the output directory" - ) - parser.add_argument( - "-n", - "--num_spliced_dataset_bins", - type=int, - default=10, - help="Number of bins for spliced dataset", - ) - parser.add_argument( - "--text_model", - type=str, - default="openai/clip-vit-base-patch32", - help="CLIP text model", - ) - parser.add_argument("--use_pooled_text", action="store_true", default=False) - args = parser.parse_args() - process_dataset( - args.captions_file, - args.video_dir, - args.output_dir, - args.num_spliced_dataset_bins, - args.text_model, - args.use_pooled_text, - ) diff --git a/scripts/inference.py b/scripts/inference.py new file mode 100644 index 00000000..900870be --- /dev/null +++ b/scripts/inference.py @@ -0,0 +1,112 @@ +import os + +import torch +import colossalai +import torch.distributed as dist +from mmengine.runner import set_random_seed + +from opensora.datasets import save_sample +from opensora.registry import MODELS, SCHEDULERS, build_module +from opensora.utils.config_utils import parse_configs +from opensora.utils.misc import to_torch_dtype +from opensora.acceleration.parallel_states import set_sequence_parallel_group +from colossalai.cluster import DistCoordinator + + +def load_prompts(prompt_path): + with open(prompt_path, "r") as f: + prompts = [line.strip() for line in f.readlines()] + return prompts + + +def main(): + # ====================================================== + # 1. cfg and init distributed env + # ====================================================== + cfg = parse_configs(training=False) + print(cfg) + + # init distributed + colossalai.launch_from_torch({}) + coordinator = DistCoordinator() + + if coordinator.world_size > 1: + set_sequence_parallel_group(dist.group.WORLD) + enable_sequence_parallelism = True + else: + enable_sequence_parallelism = False + + # ====================================================== + # 2. runtime variables + # ====================================================== + torch.set_grad_enabled(False) + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + device = "cuda" if torch.cuda.is_available() else "cpu" + dtype = to_torch_dtype(cfg.dtype) + set_random_seed(seed=cfg.seed) + prompts = load_prompts(cfg.prompt_path) + + # ====================================================== + # 3. build model & load weights + # ====================================================== + # 3.1. build model + input_size = (cfg.num_frames, *cfg.image_size) + vae = build_module(cfg.vae, MODELS) + latent_size = vae.get_latent_size(input_size) + text_encoder = build_module(cfg.text_encoder, MODELS, device=device) # T5 must be fp32 + model = build_module( + cfg.model, + MODELS, + input_size=latent_size, + in_channels=vae.out_channels, + caption_channels=text_encoder.output_dim, + model_max_length=text_encoder.model_max_length, + dtype=dtype, + enable_sequence_parallelism=enable_sequence_parallelism, + ) + text_encoder.y_embedder = model.y_embedder # hack for classifier-free guidance + + # 3.2. move to device & eval + vae = vae.to(device, dtype).eval() + model = model.to(device, dtype).eval() + + # 3.3. build scheduler + scheduler = build_module(cfg.scheduler, SCHEDULERS) + + # 3.4. support for multi-resolution + model_args = dict() + if cfg.multi_resolution: + image_size = cfg.image_size + hw = torch.tensor([image_size], device=device, dtype=dtype).repeat(cfg.batch_size, 1) + ar = torch.tensor([[image_size[0] / image_size[1]]], device=device, dtype=dtype).repeat(cfg.batch_size, 1) + model_args["data_info"] = dict(ar=ar, hw=hw) + + # ====================================================== + # 4. inference + # ====================================================== + sample_idx = 0 + save_dir = cfg.save_dir + os.makedirs(save_dir, exist_ok=True) + for i in range(0, len(prompts), cfg.batch_size): + batch_prompts = prompts[i : i + cfg.batch_size] + samples = scheduler.sample( + model, + text_encoder, + z_size=(vae.out_channels, *latent_size), + prompts=batch_prompts, + device=device, + additional_args=model_args, + ) + samples = vae.decode(samples.to(dtype)) + + if coordinator.is_master(): + for idx, sample in enumerate(samples): + print(f"Prompt: {batch_prompts[idx]}") + save_path = os.path.join(save_dir, f"sample_{sample_idx}") + save_sample(sample, fps=cfg.fps, save_path=save_path) + sample_idx += 1 + + +if __name__ == "__main__": + main() diff --git a/scripts/sample/latte/pipeline_videogen.py b/scripts/sample/latte/pipeline_videogen.py deleted file mode 100644 index 8acd6279..00000000 --- a/scripts/sample/latte/pipeline_videogen.py +++ /dev/null @@ -1,792 +0,0 @@ -# All rights reserved. -# Copyright 2024 Vchitect/Latte -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -# -# copied from https://github.com/Vchitect/Latte/blob/main/sample/pipeline_videogen.py - -import html -import inspect -import re -import urllib.parse as ul -from dataclasses import dataclass -from typing import Callable, List, Optional, Tuple, Union - -import einops -import torch -from diffusers.image_processor import VaeImageProcessor -from diffusers.models import AutoencoderKL, Transformer2DModel -from diffusers.pipelines.pipeline_utils import DiffusionPipeline -from diffusers.schedulers import DPMSolverMultistepScheduler -from diffusers.utils import ( - BACKENDS_MAPPING, - BaseOutput, - is_bs4_available, - is_ftfy_available, - logging, - replace_example_docstring, -) -from diffusers.utils.torch_utils import randn_tensor -from transformers import T5EncoderModel, T5Tokenizer - -logger = logging.get_logger(__name__) # pylint: disable=invalid-name - -if is_bs4_available(): - from bs4 import BeautifulSoup - -if is_ftfy_available(): - import ftfy - - -EXAMPLE_DOC_STRING = """ - Examples: - ```py - >>> import torch - >>> from diffusers import PixArtAlphaPipeline - - >>> # You can replace the checkpoint id with "PixArt-alpha/PixArt-XL-2-512x512" too. - >>> pipe = PixArtAlphaPipeline.from_pretrained("PixArt-alpha/PixArt-XL-2-1024-MS", torch_dtype=torch.float16) - >>> # Enable memory optimizations. - >>> pipe.enable_model_cpu_offload() - - >>> prompt = "A small cactus with a happy face in the Sahara desert." - >>> image = pipe(prompt).images[0] - ``` -""" - - -@dataclass -class VideoPipelineOutput(BaseOutput): - video: torch.Tensor - - -class VideoGenPipeline(DiffusionPipeline): - r""" - Pipeline for text-to-image generation using PixArt-Alpha. - - This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the - library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) - - Args: - vae ([`AutoencoderKL`]): - Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. - text_encoder ([`T5EncoderModel`]): - Frozen text-encoder. PixArt-Alpha uses - [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the - [t5-v1_1-xxl](https://huggingface.co/PixArt-alpha/PixArt-alpha/tree/main/t5-v1_1-xxl) variant. - tokenizer (`T5Tokenizer`): - Tokenizer of class - [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer). - transformer ([`Transformer2DModel`]): - A text conditioned `Transformer2DModel` to denoise the encoded image latents. - scheduler ([`SchedulerMixin`]): - A scheduler to be used in combination with `transformer` to denoise the encoded image latents. - """ - bad_punct_regex = re.compile( - r"[" + "#®•©™&@·º½¾¿¡§~" + "\)" + "\(" + "\]" + "\[" + "\}" + "\{" + "\|" + "\\" + "\/" + "\*" + r"]{1,}" - ) # noqa - - _optional_components = ["tokenizer", "text_encoder"] - model_cpu_offload_seq = "text_encoder->transformer->vae" - - def __init__( - self, - tokenizer: T5Tokenizer, - text_encoder: T5EncoderModel, - vae: AutoencoderKL, - transformer: Transformer2DModel, - scheduler: DPMSolverMultistepScheduler, - ): - super().__init__() - - self.register_modules( - tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler - ) - - self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) - self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) - - # Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/utils.py - def mask_text_embeddings(self, emb, mask): - if emb.shape[0] == 1: - keep_index = mask.sum().item() - return emb[:, :, :keep_index, :], keep_index # 1, 120, 4096 -> 1 7 4096 - else: - masked_feature = emb * mask[:, None, :, None] # 1 120 4096 - return masked_feature, emb.shape[2] - - # Adapted from diffusers.pipelines.deepfloyd_if.pipeline_if.encode_prompt - def encode_prompt( - self, - prompt: Union[str, List[str]], - do_classifier_free_guidance: bool = True, - negative_prompt: str = "", - num_images_per_prompt: int = 1, - device: Optional[torch.device] = None, - prompt_embeds: Optional[torch.FloatTensor] = None, - negative_prompt_embeds: Optional[torch.FloatTensor] = None, - clean_caption: bool = False, - mask_feature: bool = True, - ): - r""" - Encodes the prompt into text encoder hidden states. - - Args: - prompt (`str` or `List[str]`, *optional*): - prompt to be encoded - negative_prompt (`str` or `List[str]`, *optional*): - The prompt not to guide the image generation. If not defined, one has to pass `negative_prompt_embeds` - instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). For - PixArt-Alpha, this should be "". - do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): - whether to use classifier free guidance or not - num_images_per_prompt (`int`, *optional*, defaults to 1): - number of images that should be generated per prompt - device: (`torch.device`, *optional*): - torch device to place the resulting embeddings on - prompt_embeds (`torch.FloatTensor`, *optional*): - Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not - provided, text embeddings will be generated from `prompt` input argument. - negative_prompt_embeds (`torch.FloatTensor`, *optional*): - Pre-generated negative text embeddings. For PixArt-Alpha, it's should be the embeddings of the "" - string. - clean_caption (bool, defaults to `False`): - If `True`, the function will preprocess and clean the provided caption before encoding. - mask_feature: (bool, defaults to `True`): - If `True`, the function will mask the text embeddings. - """ - embeds_initially_provided = prompt_embeds is not None and negative_prompt_embeds is not None - - if device is None: - device = self._execution_device - - if prompt is not None and isinstance(prompt, str): - batch_size = 1 - elif prompt is not None and isinstance(prompt, list): - batch_size = len(prompt) - else: - batch_size = prompt_embeds.shape[0] - - # See Section 3.1. of the paper. - max_length = 120 - - if prompt_embeds is None: - prompt = self._text_preprocessing(prompt, clean_caption=clean_caption) - text_inputs = self.tokenizer( - prompt, - padding="max_length", - max_length=max_length, - truncation=True, - return_attention_mask=True, - add_special_tokens=True, - return_tensors="pt", - ) - text_input_ids = text_inputs.input_ids - untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids - - if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( - text_input_ids, untruncated_ids - ): - removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_length - 1 : -1]) - logger.warning( - "The following part of your input was truncated because CLIP can only handle sequences up to" - f" {max_length} tokens: {removed_text}" - ) - - attention_mask = text_inputs.attention_mask.to(device) - prompt_embeds_attention_mask = attention_mask - - prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) - prompt_embeds = prompt_embeds[0] - else: - prompt_embeds_attention_mask = torch.ones_like(prompt_embeds) - - if self.text_encoder is not None: - dtype = self.text_encoder.dtype - elif self.transformer is not None: - dtype = self.transformer.dtype - else: - dtype = None - - prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) - - bs_embed, seq_len, _ = prompt_embeds.shape - # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method - prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) - prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) - prompt_embeds_attention_mask = prompt_embeds_attention_mask.view(bs_embed, -1) - prompt_embeds_attention_mask = prompt_embeds_attention_mask.repeat(num_images_per_prompt, 1) - - # get unconditional embeddings for classifier free guidance - if do_classifier_free_guidance and negative_prompt_embeds is None: - uncond_tokens = [negative_prompt] * batch_size - uncond_tokens = self._text_preprocessing(uncond_tokens, clean_caption=clean_caption) - max_length = prompt_embeds.shape[1] - uncond_input = self.tokenizer( - uncond_tokens, - padding="max_length", - max_length=max_length, - truncation=True, - return_attention_mask=True, - add_special_tokens=True, - return_tensors="pt", - ) - attention_mask = uncond_input.attention_mask.to(device) - - negative_prompt_embeds = self.text_encoder( - uncond_input.input_ids.to(device), - attention_mask=attention_mask, - ) - negative_prompt_embeds = negative_prompt_embeds[0] - - if do_classifier_free_guidance: - # duplicate unconditional embeddings for each generation per prompt, using mps friendly method - seq_len = negative_prompt_embeds.shape[1] - - negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device) - - negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) - negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) - - # For classifier free guidance, we need to do two forward passes. - # Here we concatenate the unconditional and text embeddings into a single batch - # to avoid doing two forward passes - else: - negative_prompt_embeds = None - - # print(prompt_embeds.shape) # 1 120 4096 - # print(negative_prompt_embeds.shape) # 1 120 4096 - - # Perform additional masking. - if mask_feature and not embeds_initially_provided: - prompt_embeds = prompt_embeds.unsqueeze(1) - masked_prompt_embeds, keep_indices = self.mask_text_embeddings(prompt_embeds, prompt_embeds_attention_mask) - masked_prompt_embeds = masked_prompt_embeds.squeeze(1) - masked_negative_prompt_embeds = ( - negative_prompt_embeds[:, :keep_indices, :] if negative_prompt_embeds is not None else None - ) - - # import torch.nn.functional as F - - # padding = (0, 0, 0, 113) # (左, 右, 下, 上) - # masked_prompt_embeds_ = F.pad(masked_prompt_embeds, padding, "constant", 0) - # masked_negative_prompt_embeds_ = F.pad(masked_negative_prompt_embeds, padding, "constant", 0) - - # print(masked_prompt_embeds == masked_prompt_embeds_[:, :masked_negative_prompt_embeds.shape[1], ...]) - - return masked_prompt_embeds, masked_negative_prompt_embeds - # return masked_prompt_embeds_, masked_negative_prompt_embeds_ - - return prompt_embeds, negative_prompt_embeds - - # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs - def prepare_extra_step_kwargs(self, generator, eta): - # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature - # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. - # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 - # and should be between [0, 1] - - accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) - extra_step_kwargs = {} - if accepts_eta: - extra_step_kwargs["eta"] = eta - - # check if the scheduler accepts generator - accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) - if accepts_generator: - extra_step_kwargs["generator"] = generator - return extra_step_kwargs - - def check_inputs( - self, - prompt, - height, - width, - negative_prompt, - callback_steps, - prompt_embeds=None, - negative_prompt_embeds=None, - ): - if height % 8 != 0 or width % 8 != 0: - raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") - - if (callback_steps is None) or ( - callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) - ): - raise ValueError( - f"`callback_steps` has to be a positive integer but is {callback_steps} of type" - f" {type(callback_steps)}." - ) - - if prompt is not None and prompt_embeds is not None: - raise ValueError( - f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" - " only forward one of the two." - ) - elif prompt is None and prompt_embeds is None: - raise ValueError( - "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." - ) - elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): - raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") - - if prompt is not None and negative_prompt_embeds is not None: - raise ValueError( - f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:" - f" {negative_prompt_embeds}. Please make sure to only forward one of the two." - ) - - if negative_prompt is not None and negative_prompt_embeds is not None: - raise ValueError( - f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" - f" {negative_prompt_embeds}. Please make sure to only forward one of the two." - ) - - if prompt_embeds is not None and negative_prompt_embeds is not None: - if prompt_embeds.shape != negative_prompt_embeds.shape: - raise ValueError( - "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" - f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" - f" {negative_prompt_embeds.shape}." - ) - - # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing - def _text_preprocessing(self, text, clean_caption=False): - if clean_caption and not is_bs4_available(): - logger.warn(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`")) - logger.warn("Setting `clean_caption` to False...") - clean_caption = False - - if clean_caption and not is_ftfy_available(): - logger.warn(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`")) - logger.warn("Setting `clean_caption` to False...") - clean_caption = False - - if not isinstance(text, (tuple, list)): - text = [text] - - def process(text: str): - if clean_caption: - text = self._clean_caption(text) - text = self._clean_caption(text) - else: - text = text.lower().strip() - return text - - return [process(t) for t in text] - - # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption - def _clean_caption(self, caption): - caption = str(caption) - caption = ul.unquote_plus(caption) - caption = caption.strip().lower() - caption = re.sub("", "person", caption) - # urls: - caption = re.sub( - r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa - "", - caption, - ) # regex for urls - caption = re.sub( - r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa - "", - caption, - ) # regex for urls - # html: - caption = BeautifulSoup(caption, features="html.parser").text - - # @ - caption = re.sub(r"@[\w\d]+\b", "", caption) - - # 31C0—31EF CJK Strokes - # 31F0—31FF Katakana Phonetic Extensions - # 3200—32FF Enclosed CJK Letters and Months - # 3300—33FF CJK Compatibility - # 3400—4DBF CJK Unified Ideographs Extension A - # 4DC0—4DFF Yijing Hexagram Symbols - # 4E00—9FFF CJK Unified Ideographs - caption = re.sub(r"[\u31c0-\u31ef]+", "", caption) - caption = re.sub(r"[\u31f0-\u31ff]+", "", caption) - caption = re.sub(r"[\u3200-\u32ff]+", "", caption) - caption = re.sub(r"[\u3300-\u33ff]+", "", caption) - caption = re.sub(r"[\u3400-\u4dbf]+", "", caption) - caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption) - caption = re.sub(r"[\u4e00-\u9fff]+", "", caption) - ####################################################### - - # все виды тире / all types of dash --> "-" - caption = re.sub( - r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa - "-", - caption, - ) - - # кавычки к одному стандарту - caption = re.sub(r"[`´«»“”¨]", '"', caption) - caption = re.sub(r"[‘’]", "'", caption) - - # " - caption = re.sub(r""?", "", caption) - # & - caption = re.sub(r"&", "", caption) - - # ip adresses: - caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption) - - # article ids: - caption = re.sub(r"\d:\d\d\s+$", "", caption) - - # \n - caption = re.sub(r"\\n", " ", caption) - - # "#123" - caption = re.sub(r"#\d{1,3}\b", "", caption) - # "#12345.." - caption = re.sub(r"#\d{5,}\b", "", caption) - # "123456.." - caption = re.sub(r"\b\d{6,}\b", "", caption) - # filenames: - caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption) - - # - caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT""" - caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT""" - - caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT - caption = re.sub(r"\s+\.\s+", r" ", caption) # " . " - - # this-is-my-cute-cat / this_is_my_cute_cat - regex2 = re.compile(r"(?:\-|\_)") - if len(re.findall(regex2, caption)) > 3: - caption = re.sub(regex2, " ", caption) - - caption = ftfy.fix_text(caption) - caption = html.unescape(html.unescape(caption)) - - caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640 - caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc - caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231 - - caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption) - caption = re.sub(r"(free\s)?download(\sfree)?", "", caption) - caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption) - caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption) - caption = re.sub(r"\bpage\s+\d+\b", "", caption) - - caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a... - - caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption) - - caption = re.sub(r"\b\s+\:\s+", r": ", caption) - caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption) - caption = re.sub(r"\s+", " ", caption) - - caption.strip() - - caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption) - caption = re.sub(r"^[\'\_,\-\:;]", r"", caption) - caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption) - caption = re.sub(r"^\.\S+$", "", caption) - - return caption.strip() - - # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents - def prepare_latents( - self, batch_size, num_channels_latents, video_length, height, width, dtype, device, generator, latents=None - ): - shape = ( - batch_size, - num_channels_latents, - video_length, - height // self.vae_scale_factor, - width // self.vae_scale_factor, - ) - if isinstance(generator, list) and len(generator) != batch_size: - raise ValueError( - f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" - f" size of {batch_size}. Make sure the batch size matches the length of the generators." - ) - - if latents is None: - latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) - else: - latents = latents.to(device) - - # scale the initial noise by the standard deviation required by the scheduler - latents = latents * self.scheduler.init_noise_sigma - return latents - - @torch.no_grad() - @replace_example_docstring(EXAMPLE_DOC_STRING) - def __call__( - self, - prompt: Union[str, List[str]] = None, - negative_prompt: str = "", - num_inference_steps: int = 20, - timesteps: List[int] = None, - guidance_scale: float = 4.5, - num_images_per_prompt: Optional[int] = 1, - video_length: Optional[int] = None, - height: Optional[int] = None, - width: Optional[int] = None, - eta: float = 0.0, - generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, - latents: Optional[torch.FloatTensor] = None, - prompt_embeds: Optional[torch.FloatTensor] = None, - negative_prompt_embeds: Optional[torch.FloatTensor] = None, - output_type: Optional[str] = "pil", - return_dict: bool = True, - callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, - callback_steps: int = 1, - clean_caption: bool = True, - mask_feature: bool = True, - enable_temporal_attentions: bool = True, - enable_vae_temporal_decoder: bool = False, - ) -> Union[VideoPipelineOutput, Tuple]: - """ - Function invoked when calling the pipeline for generation. - - Args: - prompt (`str` or `List[str]`, *optional*): - The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. - instead. - negative_prompt (`str` or `List[str]`, *optional*): - The prompt or prompts not to guide the image generation. If not defined, one has to pass - `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is - less than `1`). - num_inference_steps (`int`, *optional*, defaults to 100): - The number of denoising steps. More denoising steps usually lead to a higher quality image at the - expense of slower inference. - timesteps (`List[int]`, *optional*): - Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps` - timesteps are used. Must be in descending order. - guidance_scale (`float`, *optional*, defaults to 7.0): - Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). - `guidance_scale` is defined as `w` of equation 2. of [Imagen - Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > - 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, - usually at the expense of lower image quality. - num_images_per_prompt (`int`, *optional*, defaults to 1): - The number of images to generate per prompt. - height (`int`, *optional*, defaults to self.unet.config.sample_size): - The height in pixels of the generated image. - width (`int`, *optional*, defaults to self.unet.config.sample_size): - The width in pixels of the generated image. - eta (`float`, *optional*, defaults to 0.0): - Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to - [`schedulers.DDIMScheduler`], will be ignored for others. - generator (`torch.Generator` or `List[torch.Generator]`, *optional*): - One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) - to make generation deterministic. - latents (`torch.FloatTensor`, *optional*): - Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image - generation. Can be used to tweak the same generation with different prompts. If not provided, a latents - tensor will ge generated by sampling using the supplied random `generator`. - prompt_embeds (`torch.FloatTensor`, *optional*): - Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not - provided, text embeddings will be generated from `prompt` input argument. - negative_prompt_embeds (`torch.FloatTensor`, *optional*): - Pre-generated negative text embeddings. For PixArt-Alpha this negative prompt should be "". If not - provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. - output_type (`str`, *optional*, defaults to `"pil"`): - The output format of the generate image. Choose between - [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. - return_dict (`bool`, *optional*, defaults to `True`): - Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple. - callback (`Callable`, *optional*): - A function that will be called every `callback_steps` steps during inference. The function will be - called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. - callback_steps (`int`, *optional*, defaults to 1): - The frequency at which the `callback` function will be called. If not specified, the callback will be - called at every step. - clean_caption (`bool`, *optional*, defaults to `True`): - Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to - be installed. If the dependencies are not installed, the embeddings will be created from the raw - prompt. - mask_feature (`bool` defaults to `True`): If set to `True`, the text embeddings will be masked. - - Examples: - - Returns: - [`~pipelines.ImagePipelineOutput`] or `tuple`: - If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is - returned where the first element is a list with the generated images - """ - # 1. Check inputs. Raise error if not correct - height = height or self.transformer.config.sample_size * self.vae_scale_factor - width = width or self.transformer.config.sample_size * self.vae_scale_factor - self.check_inputs(prompt, height, width, negative_prompt, callback_steps, prompt_embeds, negative_prompt_embeds) - - # 2. Default height and width to transformer - if prompt is not None and isinstance(prompt, str): - batch_size = 1 - elif prompt is not None and isinstance(prompt, list): - batch_size = len(prompt) - else: - batch_size = prompt_embeds.shape[0] - - device = self._execution_device - - # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) - # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` - # corresponds to doing no classifier free guidance. - do_classifier_free_guidance = guidance_scale > 1.0 - - # 3. Encode input prompt - prompt_embeds, negative_prompt_embeds = self.encode_prompt( - prompt, - do_classifier_free_guidance, - negative_prompt=negative_prompt, - num_images_per_prompt=num_images_per_prompt, - device=device, - prompt_embeds=prompt_embeds, - negative_prompt_embeds=negative_prompt_embeds, - clean_caption=clean_caption, - mask_feature=mask_feature, - ) - if do_classifier_free_guidance: - prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) - - # 4. Prepare timesteps - self.scheduler.set_timesteps(num_inference_steps, device=device) - timesteps = self.scheduler.timesteps - - # 5. Prepare latents. - latent_channels = self.transformer.config.in_channels - latents = self.prepare_latents( - batch_size * num_images_per_prompt, - latent_channels, - video_length, - height, - width, - prompt_embeds.dtype, - device, - generator, - latents, - ) - - # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline - extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) - - # 6.1 Prepare micro-conditions. - added_cond_kwargs = {"resolution": None, "aspect_ratio": None} - if self.transformer.config.sample_size == 128: - resolution = torch.tensor([height, width]).repeat(batch_size * num_images_per_prompt, 1) - aspect_ratio = torch.tensor([float(height / width)]).repeat(batch_size * num_images_per_prompt, 1) - resolution = resolution.to(dtype=prompt_embeds.dtype, device=device) - aspect_ratio = aspect_ratio.to(dtype=prompt_embeds.dtype, device=device) - added_cond_kwargs = {"resolution": resolution, "aspect_ratio": aspect_ratio} - - # 7. Denoising loop - num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) - - with self.progress_bar(total=num_inference_steps) as progress_bar: - for i, t in enumerate(timesteps): - latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents - latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) - - current_timestep = t - if not torch.is_tensor(current_timestep): - # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can - # This would be a good case for the `match` statement (Python 3.10+) - is_mps = latent_model_input.device.type == "mps" - if isinstance(current_timestep, float): - dtype = torch.float32 if is_mps else torch.float64 - else: - dtype = torch.int32 if is_mps else torch.int64 - current_timestep = torch.tensor([current_timestep], dtype=dtype, device=latent_model_input.device) - elif len(current_timestep.shape) == 0: - current_timestep = current_timestep[None].to(latent_model_input.device) - # broadcast to batch dimension in a way that's compatible with ONNX/Core ML - current_timestep = current_timestep.expand(latent_model_input.shape[0]) - - # predict noise model_output - noise_pred = self.transformer( - latent_model_input, - encoder_hidden_states=prompt_embeds, - timestep=current_timestep, - added_cond_kwargs=added_cond_kwargs, - enable_temporal_attentions=enable_temporal_attentions, - return_dict=False, - )[0] - - # perform guidance - if do_classifier_free_guidance: - noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) - noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) - - # learned sigma - if self.transformer.config.out_channels // 2 == latent_channels: - noise_pred = noise_pred.chunk(2, dim=1)[0] - else: - noise_pred = noise_pred - - # compute previous image: x_t -> x_t-1 - latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] - - # call the callback, if provided - if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): - progress_bar.update() - if callback is not None and i % callback_steps == 0: - step_idx = i // getattr(self.scheduler, "order", 1) - callback(step_idx, t, latents) - - if not output_type == "latents": - if enable_vae_temporal_decoder: - video = self.decode_latents_with_temporal_decoder(latents) - else: - video = self.decode_latents(latents) - else: - video = latents - return VideoPipelineOutput(video=video) - - # Offload all models - self.maybe_free_model_hooks() - - if not return_dict: - return (video,) - - return VideoPipelineOutput(video=video) - - def decode_latents(self, latents): - video_length = latents.shape[2] - latents = 1 / self.vae.config.scaling_factor * latents - latents = einops.rearrange(latents, "b c f h w -> (b f) c h w") - video = [] - for frame_idx in range(latents.shape[0]): - video.append(self.vae.decode(latents[frame_idx : frame_idx + 1]).sample) - video = torch.cat(video) - video = einops.rearrange(video, "(b f) c h w -> b f h w c", f=video_length) - video = ((video / 2.0 + 0.5).clamp(0, 1) * 255).to(dtype=torch.uint8).cpu().contiguous() - # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16 - return video - - def decode_latents_with_temporal_decoder(self, latents): - video_length = latents.shape[2] - latents = 1 / self.vae.config.scaling_factor * latents - latents = einops.rearrange(latents, "b c f h w -> (b f) c h w") - video = [] - - decode_chunk_size = 14 - for frame_idx in range(0, latents.shape[0], decode_chunk_size): - num_frames_in = latents[frame_idx : frame_idx + decode_chunk_size].shape[0] - - decode_kwargs = {} - decode_kwargs["num_frames"] = num_frames_in - - video.append(self.vae.decode(latents[frame_idx : frame_idx + decode_chunk_size], **decode_kwargs).sample) - - video = torch.cat(video) - video = einops.rearrange(video, "(b f) c h w -> b f h w c", f=video_length) - video = ((video / 2.0 + 0.5).clamp(0, 1) * 255).to(dtype=torch.uint8).cpu().contiguous() - # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16 - return video diff --git a/scripts/sample/latte/sample.sh b/scripts/sample/latte/sample.sh deleted file mode 100644 index c85272f1..00000000 --- a/scripts/sample/latte/sample.sh +++ /dev/null @@ -1,19 +0,0 @@ -#!/usr/bin/env bash - -# get args -GPUS=${1:-8} - -# get root dir -FOLDER_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )" -ROOT_DIR=$FOLDER_DIR/../../.. - -# go to root dir -cd $ROOT_DIR - - -export PYTHONPATH=$FOLDER_DIR:$PYTHONPATH -python $FOLDER_DIR/sample_t2v.py \ - --checkpoint /home/lishenggui/projects/sora/hf-weights/models--maxin-cn--Latte/snapshots/8f0591220fa329f9d917086810b3c0f6544a87c7/t2v.pt \ - --model_path /home/lishenggui/projects/sora/hf-weights/models--maxin-cn--Latte/snapshots/8f0591220fa329f9d917086810b3c0f6544a87c7/t2v_required_models/ \ - --text_prompt "A dog in astronaut suit and sunglasses floating in space" \ - --output_path $ROOT_DIR/outputs/latte diff --git a/scripts/sample/latte/sample_t2v.py b/scripts/sample/latte/sample_t2v.py deleted file mode 100644 index c65d0073..00000000 --- a/scripts/sample/latte/sample_t2v.py +++ /dev/null @@ -1,242 +0,0 @@ -# All rights reserved. -# Copyright 2024 Vchitect/Latte -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -# -# modified from https://github.com/Vchitect/Latte/blob/main/sample/sample_t2v.py - -import argparse -import os -import sys - -import torch -from diffusers.models import AutoencoderKL, AutoencoderKLTemporalDecoder -from diffusers.schedulers import ( - DDIMScheduler, - DDPMScheduler, - DEISMultistepScheduler, - DPMSolverMultistepScheduler, - EulerAncestralDiscreteScheduler, - EulerDiscreteScheduler, - HeunDiscreteScheduler, - KDPM2AncestralDiscreteScheduler, - PNDMScheduler, -) -from diffusers.schedulers.scheduling_dpmsolver_singlestep import DPMSolverSinglestepScheduler -from transformers import T5EncoderModel, T5Tokenizer - -sys.path.append(os.path.split(sys.path[0])[0]) -import imageio -from pipeline_videogen import VideoGenPipeline -from utils import save_video_grid - -from download import find_model -from open_sora.modeling import LatteT2V - - -def parse_args(): - parser = argparse.ArgumentParser() - parser.add_argument("--model_path", type=str, required=True, help="The path to the pretrained model files") - parser.add_argument("--checkpoint", type=str, required=True, help="The path to the t2v.pt file.") - parser.add_argument("--output_path", type=str, required=True, help="The path to save the output") - - # generation configs - parser.add_argument( - "--text_prompt", type=str, nargs="+", required=True, help="The text prompt to generate the video." - ) - parser.add_argument("--video_length", type=int, default=16, help="The number of frames in the video.") - parser.add_argument("--image_height", type=int, default=256, help="The size of the generated images.") - parser.add_argument("--image_width", type=int, default=256, help="The size of the generated images.") - parser.add_argument("--guidance_scale", type=float, default=7.5, help="The scale of the guidance loss.") - parser.add_argument("--sample_method", type=str, default="PNDM", help="The sampling method to use.") - parser.add_argument("--num_sampling_steps", type=int, default=50, help="The number of sampling steps.") - parser.add_argument( - "--enable_temporal_attentions", action="store_true", default=True, help="Whether to enable temporal attentions." - ) - parser.add_argument( - "--enable_vae_temporal_decoder", - action="store_true", - default=True, - help="Whether to enable the VAE temporal decoder.", - ) - - # Scheduler configs - parser.add_argument("--beta_start", type=float, default=0.0001) - parser.add_argument("--beta_end", type=float, default=0.02) - parser.add_argument("--beta_schedule", type=str, default="linear") - parser.add_argument("--variance_type", type=str, default="learned_range") - - args = parser.parse_args() - return args - - -def main(args): - torch.set_grad_enabled(False) - device = "cuda" if torch.cuda.is_available() else "cpu" - - transformer_model = LatteT2V.from_pretrained_2d( - args.model_path, subfolder="transformer", video_length=args.video_length - ).to(device, dtype=torch.float16) - state_dict = find_model(args.checkpoint) - transformer_model.load_state_dict(state_dict["model"]) - - if args.enable_vae_temporal_decoder: - vae = AutoencoderKLTemporalDecoder.from_pretrained( - args.model_path, subfolder="vae_temporal_decoder", torch_dtype=torch.float16 - ).to(device) - else: - vae = AutoencoderKL.from_pretrained(args.model_path, subfolder="vae", torch_dtype=torch.float16).to(device) - tokenizer = T5Tokenizer.from_pretrained(args.model_path, subfolder="tokenizer") - text_encoder = T5EncoderModel.from_pretrained( - args.model_path, subfolder="text_encoder", torch_dtype=torch.float16 - ).to(device) - - # set eval mode - transformer_model.eval() - vae.eval() - text_encoder.eval() - - if args.sample_method == "DDIM": - scheduler = DDIMScheduler.from_pretrained( - args.model_path, - subfolder="scheduler", - beta_start=args.beta_start, - beta_end=args.beta_end, - beta_schedule=args.beta_schedule, - variance_type=args.variance_type, - ) - elif args.sample_method == "EulerDiscrete": - scheduler = EulerDiscreteScheduler.from_pretrained( - args.model_path, - subfolder="scheduler", - beta_start=args.beta_start, - beta_end=args.beta_end, - beta_schedule=args.beta_schedule, - variance_type=args.variance_type, - ) - elif args.sample_method == "DDPM": - scheduler = DDPMScheduler.from_pretrained( - args.model_path, - subfolder="scheduler", - beta_start=args.beta_start, - beta_end=args.beta_end, - beta_schedule=args.beta_schedule, - variance_type=args.variance_type, - ) - elif args.sample_method == "DPMSolverMultistep": - scheduler = DPMSolverMultistepScheduler.from_pretrained( - args.model_path, - subfolder="scheduler", - beta_start=args.beta_start, - beta_end=args.beta_end, - beta_schedule=args.beta_schedule, - variance_type=args.variance_type, - ) - elif args.sample_method == "DPMSolverSinglestep": - scheduler = DPMSolverSinglestepScheduler.from_pretrained( - args.model_path, - subfolder="scheduler", - beta_start=args.beta_start, - beta_end=args.beta_end, - beta_schedule=args.beta_schedule, - variance_type=args.variance_type, - ) - elif args.sample_method == "PNDM": - scheduler = PNDMScheduler.from_pretrained( - args.model_path, - subfolder="scheduler", - beta_start=args.beta_start, - beta_end=args.beta_end, - beta_schedule=args.beta_schedule, - variance_type=args.variance_type, - ) - elif args.sample_method == "HeunDiscrete": - scheduler = HeunDiscreteScheduler.from_pretrained( - args.model_path, - subfolder="scheduler", - beta_start=args.beta_start, - beta_end=args.beta_end, - beta_schedule=args.beta_schedule, - variance_type=args.variance_type, - ) - elif args.sample_method == "EulerAncestralDiscrete": - scheduler = EulerAncestralDiscreteScheduler.from_pretrained( - args.model_path, - subfolder="scheduler", - beta_start=args.beta_start, - beta_end=args.beta_end, - beta_schedule=args.beta_schedule, - variance_type=args.variance_type, - ) - elif args.sample_method == "DEISMultistep": - scheduler = DEISMultistepScheduler.from_pretrained( - args.model_path, - subfolder="scheduler", - beta_start=args.beta_start, - beta_end=args.beta_end, - beta_schedule=args.beta_schedule, - variance_type=args.variance_type, - ) - elif args.sample_method == "KDPM2AncestralDiscrete": - scheduler = KDPM2AncestralDiscreteScheduler.from_pretrained( - args.model_path, - subfolder="scheduler", - beta_start=args.beta_start, - beta_end=args.beta_end, - beta_schedule=args.beta_schedule, - variance_type=args.variance_type, - ) - - videogen_pipeline = VideoGenPipeline( - vae=vae, text_encoder=text_encoder, tokenizer=tokenizer, scheduler=scheduler, transformer=transformer_model - ).to(device) - # videogen_pipeline.enable_xformers_memory_efficient_attention() - - if not os.path.exists(args.output_path): - os.makedirs(args.output_path, exist_ok=True) - - video_grids = [] - for prompt in args.text_prompt: - print("Processing the ({}) prompt".format(prompt)) - videos = videogen_pipeline( - prompt, - video_length=args.video_length, - height=args.image_height, - width=args.image_width, - num_inference_steps=args.num_sampling_steps, - guidance_scale=args.guidance_scale, - enable_temporal_attentions=args.enable_temporal_attentions, - num_images_per_prompt=1, - mask_feature=True, - enable_vae_temporal_decoder=args.enable_vae_temporal_decoder, - ).video - try: - save_path = os.path.join(args.output_path, prompt.replace(" ", "_") + "_webv-imageio.mp4") - imageio.mimwrite(save_path, videos[0], fps=8, quality=9) # highest quality is 10, lowest is 0 - except: - print("Error when saving {}".format(prompt)) - video_grids.append(videos) - video_grids = torch.cat(video_grids, dim=0) - - video_grids = save_video_grid(video_grids) - - # torchvision.io.write_video(args.output_path + '_%04d' % args.run_time + '-.mp4', video_grids, fps=6) - save_path = os.path.join(args.output_path, "grid.mp4") - imageio.mimwrite(save_path, video_grids, fps=8, quality=5) - print("save path {}".format(abspath(args.output_path))) - - # save_videos_grid(video, f"./{prompt}.gif") - - -if __name__ == "__main__": - args = parse_args() - main(args) diff --git a/scripts/sample/latte/utils.py b/scripts/sample/latte/utils.py deleted file mode 100644 index 3a9e4f12..00000000 --- a/scripts/sample/latte/utils.py +++ /dev/null @@ -1,471 +0,0 @@ -# All rights reserved. -# Copyright 2024 Vchitect/Latte -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -# -# copied from https://github.com/Vchitect/Latte/blob/main/utils.py - -import html -import logging -import math -import os -import re -import subprocess -import urllib.parse as ul -from collections import OrderedDict -from typing import Iterable, Union - -import torch -import torch.distributed as dist -from diffusers.utils import is_bs4_available, is_ftfy_available - -# from torch._six import inf -from torch import inf -from torch.utils.tensorboard import SummaryWriter - -if is_bs4_available(): - from bs4 import BeautifulSoup - -if is_ftfy_available(): - import ftfy - -_tensor_or_tensors = Union[torch.Tensor, Iterable[torch.Tensor]] - - -################################################################################# -# Training Clip Gradients # -################################################################################# - - -def get_grad_norm(parameters: _tensor_or_tensors, norm_type: float = 2.0) -> torch.Tensor: - r""" - Copy from torch.nn.utils.clip_grad_norm_ - - Clips gradient norm of an iterable of parameters. - - The norm is computed over all gradients together, as if they were - concatenated into a single vector. Gradients are modified in-place. - - Args: - parameters (Iterable[Tensor] or Tensor): an iterable of Tensors or a - single Tensor that will have gradients normalized - max_norm (float or int): max norm of the gradients - norm_type (float or int): type of the used p-norm. Can be ``'inf'`` for - infinity norm. - error_if_nonfinite (bool): if True, an error is thrown if the total - norm of the gradients from :attr:`parameters` is ``nan``, - ``inf``, or ``-inf``. Default: False (will switch to True in the future) - - Returns: - Total norm of the parameter gradients (viewed as a single vector). - """ - if isinstance(parameters, torch.Tensor): - parameters = [parameters] - grads = [p.grad for p in parameters if p.grad is not None] - norm_type = float(norm_type) - if len(grads) == 0: - return torch.tensor(0.0) - device = grads[0].device - if norm_type == inf: - norms = [g.detach().abs().max().to(device) for g in grads] - total_norm = norms[0] if len(norms) == 1 else torch.max(torch.stack(norms)) - else: - total_norm = torch.norm(torch.stack([torch.norm(g.detach(), norm_type).to(device) for g in grads]), norm_type) - return total_norm - - -def clip_grad_norm_( - parameters: _tensor_or_tensors, - max_norm: float, - norm_type: float = 2.0, - error_if_nonfinite: bool = False, - clip_grad=True, -) -> torch.Tensor: - r""" - Copy from torch.nn.utils.clip_grad_norm_ - - Clips gradient norm of an iterable of parameters. - - The norm is computed over all gradients together, as if they were - concatenated into a single vector. Gradients are modified in-place. - - Args: - parameters (Iterable[Tensor] or Tensor): an iterable of Tensors or a - single Tensor that will have gradients normalized - max_norm (float or int): max norm of the gradients - norm_type (float or int): type of the used p-norm. Can be ``'inf'`` for - infinity norm. - error_if_nonfinite (bool): if True, an error is thrown if the total - norm of the gradients from :attr:`parameters` is ``nan``, - ``inf``, or ``-inf``. Default: False (will switch to True in the future) - - Returns: - Total norm of the parameter gradients (viewed as a single vector). - """ - if isinstance(parameters, torch.Tensor): - parameters = [parameters] - grads = [p.grad for p in parameters if p.grad is not None] - max_norm = float(max_norm) - norm_type = float(norm_type) - if len(grads) == 0: - return torch.tensor(0.0) - device = grads[0].device - if norm_type == inf: - norms = [g.detach().abs().max().to(device) for g in grads] - total_norm = norms[0] if len(norms) == 1 else torch.max(torch.stack(norms)) - else: - total_norm = torch.norm(torch.stack([torch.norm(g.detach(), norm_type).to(device) for g in grads]), norm_type) - # print(total_norm) - - if clip_grad: - if error_if_nonfinite and torch.logical_or(total_norm.isnan(), total_norm.isinf()): - raise RuntimeError( - f"The total norm of order {norm_type} for gradients from " - "`parameters` is non-finite, so it cannot be clipped. To disable " - "this error and scale the gradients by the non-finite norm anyway, " - "set `error_if_nonfinite=False`" - ) - clip_coef = max_norm / (total_norm + 1e-6) - # Note: multiplying by the clamped coef is redundant when the coef is clamped to 1, but doing so - # avoids a `if clip_coef < 1:` conditional which can require a CPU <=> device synchronization - # when the gradients do not reside in CPU memory. - clip_coef_clamped = torch.clamp(clip_coef, max=1.0) - for g in grads: - g.detach().mul_(clip_coef_clamped.to(g.device)) - # gradient_cliped = torch.norm(torch.stack([torch.norm(g.detach(), norm_type).to(device) for g in grads]), norm_type) - # print(gradient_cliped) - return total_norm - - -def get_experiment_dir(root_dir, args): - # if args.pretrained is not None and 'Latte-XL-2-256x256.pt' not in args.pretrained: - # root_dir += '-WOPRE' - if args.use_compile: - root_dir += "-Compile" # speedup by torch compile - if args.fixed_spatial: - root_dir += "-FixedSpa" - if args.enable_xformers_memory_efficient_attention: - root_dir += "-Xfor" - if args.gradient_checkpointing: - root_dir += "-Gc" - if args.mixed_precision: - root_dir += "-Amp" - if args.image_size == 512: - root_dir += "-512" - return root_dir - - -################################################################################# -# Training Logger # -################################################################################# - - -def create_logger(logging_dir): - """ - Create a logger that writes to a log file and stdout. - """ - if dist.get_rank() == 0: # real logger - logging.basicConfig( - level=logging.INFO, - # format='[\033[34m%(asctime)s\033[0m] %(message)s', - format="[%(asctime)s] %(message)s", - datefmt="%Y-%m-%d %H:%M:%S", - handlers=[logging.StreamHandler(), logging.FileHandler(f"{logging_dir}/log.txt")], - ) - logger = logging.getLogger(__name__) - - else: # dummy logger (does nothing) - logger = logging.getLogger(__name__) - logger.addHandler(logging.NullHandler()) - return logger - - -def create_tensorboard(tensorboard_dir): - """ - Create a tensorboard that saves losses. - """ - if dist.get_rank() == 0: # real tensorboard - # tensorboard - writer = SummaryWriter(tensorboard_dir) - - return writer - - -def write_tensorboard(writer, *args): - """ - write the loss information to a tensorboard file. - Only for pytorch DDP mode. - """ - if dist.get_rank() == 0: # real tensorboard - writer.add_scalar(args[0], args[1], args[2]) - - -################################################################################# -# EMA Update/ DDP Training Utils # -################################################################################# - - -@torch.no_grad() -def update_ema(ema_model, model, decay=0.9999): - """ - Step the EMA model towards the current model. - """ - ema_params = OrderedDict(ema_model.named_parameters()) - model_params = OrderedDict(model.named_parameters()) - - for name, param in model_params.items(): - # TODO: Consider applying only to params that require_grad to avoid small numerical changes of pos_embed - ema_params[name].mul_(decay).add_(param.data, alpha=1 - decay) - - -def requires_grad(model, flag=True): - """ - Set requires_grad flag for all parameters in a model. - """ - for p in model.parameters(): - p.requires_grad = flag - - -def cleanup(): - """ - End DDP training. - """ - dist.destroy_process_group() - - -def setup_distributed(backend="nccl", port=None): - """Initialize distributed training environment. - support both slurm and torch.distributed.launch - see torch.distributed.init_process_group() for more details - """ - num_gpus = torch.cuda.device_count() - - if "SLURM_JOB_ID" in os.environ: - rank = int(os.environ["SLURM_PROCID"]) - world_size = int(os.environ["SLURM_NTASKS"]) - node_list = os.environ["SLURM_NODELIST"] - addr = subprocess.getoutput(f"scontrol show hostname {node_list} | head -n1") - # specify master port - if port is not None: - os.environ["MASTER_PORT"] = str(port) - elif "MASTER_PORT" not in os.environ: - # os.environ["MASTER_PORT"] = "29566" - os.environ["MASTER_PORT"] = str(29567 + num_gpus) - if "MASTER_ADDR" not in os.environ: - os.environ["MASTER_ADDR"] = addr - os.environ["WORLD_SIZE"] = str(world_size) - os.environ["LOCAL_RANK"] = str(rank % num_gpus) - os.environ["RANK"] = str(rank) - else: - rank = int(os.environ["RANK"]) - world_size = int(os.environ["WORLD_SIZE"]) - - # torch.cuda.set_device(rank % num_gpus) - - dist.init_process_group( - backend=backend, - world_size=world_size, - rank=rank, - ) - - -################################################################################# -# Testing Utils # -################################################################################# - - -def save_video_grid(video, nrow=None): - b, t, h, w, c = video.shape - - if nrow is None: - nrow = math.ceil(math.sqrt(b)) - ncol = math.ceil(b / nrow) - padding = 1 - video_grid = torch.zeros((t, (padding + h) * nrow + padding, (padding + w) * ncol + padding, c), dtype=torch.uint8) - - print(video_grid.shape) - for i in range(b): - r = i // ncol - c = i % ncol - start_r = (padding + h) * r - start_c = (padding + w) * c - video_grid[:, start_r : start_r + h, start_c : start_c + w] = video[i] - - return video_grid - - -################################################################################# -# MMCV Utils # -################################################################################# - - -def collect_env(): - # Copyright (c) OpenMMLab. All rights reserved. - from mmcv.utils import collect_env as collect_base_env - from mmcv.utils import get_git_hash - - """Collect the information of the running environments.""" - - env_info = collect_base_env() - env_info["MMClassification"] = get_git_hash()[:7] - - for name, val in env_info.items(): - print(f"{name}: {val}") - - print(torch.cuda.get_arch_list()) - print(torch.version.cuda) - - -################################################################################# -# Pixart-alpha Utils # -################################################################################# - -bad_punct_regex = re.compile( - r"[" + "#®•©™&@·º½¾¿¡§~" + "\)" + "\(" + "\]" + "\[" + "\}" + "\{" + "\|" + "\\" + "\/" + "\*" + r"]{1,}" -) - - -def text_preprocessing(text, clean_caption=False): - if clean_caption and not is_bs4_available(): - clean_caption = False - - if clean_caption and not is_ftfy_available(): - clean_caption = False - - if not isinstance(text, (tuple, list)): - text = [text] - - def process(text: str): - if clean_caption: - text = clean_caption(text) - text = clean_caption(text) - else: - text = text.lower().strip() - return text - - return [process(t) for t in text] - - -# Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption -def clean_caption(caption): - caption = str(caption) - caption = ul.unquote_plus(caption) - caption = caption.strip().lower() - caption = re.sub("", "person", caption) - # urls: - caption = re.sub( - r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa - "", - caption, - ) # regex for urls - caption = re.sub( - r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa - "", - caption, - ) # regex for urls - # html: - caption = BeautifulSoup(caption, features="html.parser").text - - # @ - caption = re.sub(r"@[\w\d]+\b", "", caption) - - # 31C0—31EF CJK Strokes - # 31F0—31FF Katakana Phonetic Extensions - # 3200—32FF Enclosed CJK Letters and Months - # 3300—33FF CJK Compatibility - # 3400—4DBF CJK Unified Ideographs Extension A - # 4DC0—4DFF Yijing Hexagram Symbols - # 4E00—9FFF CJK Unified Ideographs - caption = re.sub(r"[\u31c0-\u31ef]+", "", caption) - caption = re.sub(r"[\u31f0-\u31ff]+", "", caption) - caption = re.sub(r"[\u3200-\u32ff]+", "", caption) - caption = re.sub(r"[\u3300-\u33ff]+", "", caption) - caption = re.sub(r"[\u3400-\u4dbf]+", "", caption) - caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption) - caption = re.sub(r"[\u4e00-\u9fff]+", "", caption) - ####################################################### - - # все виды тире / all types of dash --> "-" - caption = re.sub( - r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa - "-", - caption, - ) - - # кавычки к одному стандарту - caption = re.sub(r"[`´«»“”¨]", '"', caption) - caption = re.sub(r"[‘’]", "'", caption) - - # " - caption = re.sub(r""?", "", caption) - # & - caption = re.sub(r"&", "", caption) - - # ip adresses: - caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption) - - # article ids: - caption = re.sub(r"\d:\d\d\s+$", "", caption) - - # \n - caption = re.sub(r"\\n", " ", caption) - - # "#123" - caption = re.sub(r"#\d{1,3}\b", "", caption) - # "#12345.." - caption = re.sub(r"#\d{5,}\b", "", caption) - # "123456.." - caption = re.sub(r"\b\d{6,}\b", "", caption) - # filenames: - caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption) - - # - caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT""" - caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT""" - - caption = re.sub(bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT - caption = re.sub(r"\s+\.\s+", r" ", caption) # " . " - - # this-is-my-cute-cat / this_is_my_cute_cat - regex2 = re.compile(r"(?:\-|\_)") - if len(re.findall(regex2, caption)) > 3: - caption = re.sub(regex2, " ", caption) - - caption = ftfy.fix_text(caption) - caption = html.unescape(html.unescape(caption)) - - caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640 - caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc - caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231 - - caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption) - caption = re.sub(r"(free\s)?download(\sfree)?", "", caption) - caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption) - caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption) - caption = re.sub(r"\bpage\s+\d+\b", "", caption) - - caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a... - - caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption) - - caption = re.sub(r"\b\s+\:\s+", r": ", caption) - caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption) - caption = re.sub(r"\s+", " ", caption) - - caption.strip() - - caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption) - caption = re.sub(r"^[\'\_,\-\:;]", r"", caption) - caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption) - caption = re.sub(r"^\.\S+$", "", caption) - - return caption.strip() diff --git a/scripts/train.py b/scripts/train.py new file mode 100644 index 00000000..9f611b7d --- /dev/null +++ b/scripts/train.py @@ -0,0 +1,287 @@ +from copy import deepcopy + +import colossalai +import torch +import torch.distributed as dist +import wandb +from colossalai.booster import Booster +from colossalai.booster.plugin import LowLevelZeroPlugin +from colossalai.cluster import DistCoordinator +from colossalai.nn.optimizer import HybridAdam +from colossalai.utils import get_current_device +from tqdm import tqdm + +from opensora.acceleration.checkpoint import set_grad_checkpoint +from opensora.acceleration.parallel_states import ( + get_data_parallel_group, + set_data_parallel_group, + set_sequence_parallel_group, +) +from opensora.acceleration.plugin import ZeroSeqParallelPlugin +from opensora.datasets import DatasetFromCSV, get_transforms_image, get_transforms_video, prepare_dataloader +from opensora.registry import MODELS, SCHEDULERS, build_module +from opensora.utils.ckpt_utils import create_logger, load, model_sharding, record_model_param_shape, save +from opensora.utils.config_utils import ( + create_experiment_workspace, + create_tensorboard_writer, + parse_configs, + save_training_config, +) +from opensora.utils.misc import all_reduce_mean, format_numel_str, get_model_numel, requires_grad, to_torch_dtype +from opensora.utils.train_utils import update_ema + + +def main(): + # ====================================================== + # 1. args & cfg + # ====================================================== + cfg = parse_configs(training=True) + print(cfg) + exp_name, exp_dir = create_experiment_workspace(cfg) + save_training_config(cfg._cfg_dict, exp_dir) + + # ====================================================== + # 2. runtime variables & colossalai launch + # ====================================================== + assert torch.cuda.is_available(), "Training currently requires at least one GPU." + assert cfg.dtype in ["fp16", "bf16"], f"Unknown mixed precision {cfg.dtype}" + + # 2.1. colossalai init distributed training + colossalai.launch_from_torch({}) + coordinator = DistCoordinator() + device = get_current_device() + dtype = to_torch_dtype(cfg.dtype) + + # 2.2. init logger, tensorboard & wandb + if not coordinator.is_master(): + logger = create_logger(None) + else: + logger = create_logger(exp_dir) + logger.info(f"Experiment directory created at {exp_dir}") + + writer = create_tensorboard_writer(exp_dir) + if cfg.wandb: + wandb.init(project="minisora", name=exp_name, config=cfg._cfg_dict) + + # 2.3. initialize ColossalAI booster + if cfg.plugin == "zero2": + plugin = LowLevelZeroPlugin( + stage=2, + precision=cfg.dtype, + initial_scale=2**16, + max_norm=cfg.grad_clip, + ) + set_data_parallel_group(dist.group.WORLD) + elif cfg.plugin == "zero2-seq": + plugin = ZeroSeqParallelPlugin( + sp_size=cfg.sp_size, + stage=2, + precision=cfg.dtype, + initial_scale=2**16, + max_norm=cfg.grad_clip, + ) + set_sequence_parallel_group(plugin.sp_group) + set_data_parallel_group(plugin.dp_group) + else: + raise ValueError(f"Unknown plugin {cfg.plugin}") + booster = Booster(plugin=plugin) + + # ====================================================== + # 3. build dataset and dataloader + # ====================================================== + dataset = DatasetFromCSV( + cfg.data_path, + # TODO: change transforms + transform=( + get_transforms_video(cfg.image_size[0]) + if not cfg.use_image_transform + else get_transforms_image(cfg.image_size[0]) + ), + num_frames=cfg.num_frames, + frame_interval=cfg.frame_interval, + root=cfg.root, + ) + + # TODO: use plugin's prepare dataloader + # a batch contains: + # { + # "video": torch.Tensor, # [B, C, T, H, W], + # "text": List[str], + # } + dataloader = prepare_dataloader( + dataset, + batch_size=cfg.batch_size, + num_workers=cfg.num_workers, + shuffle=True, + drop_last=True, + pin_memory=True, + process_group=get_data_parallel_group(), + ) + logger.info(f"Dataset contains {len(dataset):,} videos ({cfg.data_path})") + + total_batch_size = cfg.batch_size * dist.get_world_size() // cfg.sp_size + logger.info(f"Total batch size: {total_batch_size}") + + # ====================================================== + # 4. build model + # ====================================================== + # 4.1. build model + input_size = (cfg.num_frames, *cfg.image_size) + vae = build_module(cfg.vae, MODELS) + latent_size = vae.get_latent_size(input_size) + text_encoder = build_module(cfg.text_encoder, MODELS, device=device) # T5 must be fp32 + model = build_module( + cfg.model, + MODELS, + input_size=latent_size, + in_channels=vae.out_channels, + caption_channels=text_encoder.output_dim, + model_max_length=text_encoder.model_max_length, + dtype=dtype, + ) + model_numel, model_numel_trainable = get_model_numel(model) + logger.info( + f"Trainable model params: {format_numel_str(model_numel_trainable)}, Total model params: {format_numel_str(model_numel)}" + ) + + # 4.2. create ema + ema = deepcopy(model).to(torch.float32).to(device) + requires_grad(ema, False) + ema_shape_dict = record_model_param_shape(ema) + + # 4.3. move to device + vae = vae.to(device, dtype) + model = model.to(device, dtype) + + # 4.4. build scheduler + scheduler = build_module(cfg.scheduler, SCHEDULERS) + + # 4.5. setup optimizer + optimizer = HybridAdam( + filter(lambda p: p.requires_grad, model.parameters()), lr=cfg.lr, weight_decay=0, adamw_mode=True + ) + lr_scheduler = None + + # 4.6. prepare for training + if cfg.grad_checkpoint: + set_grad_checkpoint(model) + model.train() + update_ema(ema, model, decay=0, sharded=False) + ema.eval() + + # ======================================================= + # 5. boost model for distributed training with colossalai + # ======================================================= + torch.set_default_dtype(dtype) + model, optimizer, _, dataloader, lr_scheduler = booster.boost( + model=model, optimizer=optimizer, lr_scheduler=lr_scheduler, dataloader=dataloader + ) + torch.set_default_dtype(torch.float) + num_steps_per_epoch = len(dataloader) + logger.info("Boost model for distributed training") + + # ======================================================= + # 6. training loop + # ======================================================= + start_epoch = start_step = log_step = sampler_start_idx = 0 + running_loss = 0.0 + + # 6.1. resume training + if cfg.load is not None: + logger.info("Loading checkpoint") + start_epoch, start_step, sampler_start_idx = load(booster, model, ema, optimizer, lr_scheduler, cfg.load) + logger.info(f"Loaded checkpoint {cfg.load} at epoch {start_epoch} step {start_step}") + logger.info(f"Training for {cfg.epochs} epochs with {num_steps_per_epoch} steps per epoch") + + dataloader.sampler.set_start_index(sampler_start_idx) + model_sharding(ema) + + # 6.2. training loop + for epoch in range(start_epoch, cfg.epochs): + dataloader.sampler.set_epoch(epoch) + dataloader_iter = iter(dataloader) + logger.info(f"Beginning epoch {epoch}...") + + with tqdm( + range(start_step, num_steps_per_epoch), + desc=f"Epoch {epoch}", + disable=not coordinator.is_master(), + total=num_steps_per_epoch, + initial=start_step, + ) as pbar: + for step in pbar: + batch = next(dataloader_iter) + x = batch["video"].to(device, dtype) # [B, C, T, H, W] + y = batch["text"] + + with torch.no_grad(): + # Prepare visual inputs + x = vae.encode(x) # [B, C, T, H/P, W/P] + # Prepare text inputs + model_args = text_encoder.encode(y) + + # Diffusion + t = torch.randint(0, scheduler.num_timesteps, (x.shape[0],), device=device) + loss_dict = scheduler.training_losses(model, x, t, model_args) + + # Backward & update + loss = loss_dict["loss"].mean() + booster.backward(loss=loss, optimizer=optimizer) + optimizer.step() + optimizer.zero_grad() + + # Update EMA + update_ema(ema, model.module, optimizer=optimizer) + + # Log loss values: + all_reduce_mean(loss) + running_loss += loss.item() + global_step = epoch * num_steps_per_epoch + step + log_step += 1 + + # Log to tensorboard + if coordinator.is_master() and (global_step + 1) % cfg.log_every == 0: + avg_loss = running_loss / log_step + pbar.set_postfix({"loss": avg_loss, "step": step, "global_step": global_step}) + running_loss = 0 + log_step = 0 + writer.add_scalar("loss", loss.item(), global_step) + if cfg.wandb: + wandb.log( + { + "iter": global_step, + "num_samples": global_step * total_batch_size, + "epoch": epoch, + "loss": loss.item(), + "avg_loss": avg_loss, + }, + step=global_step, + ) + + # Save checkpoint + if cfg.ckpt_every > 0 and (global_step + 1) % cfg.ckpt_every == 0: + save( + booster, + model, + ema, + optimizer, + lr_scheduler, + epoch, + step + 1, + global_step + 1, + cfg.batch_size, + coordinator, + exp_dir, + ema_shape_dict, + ) + logger.info( + f"Saved checkpoint at epoch {epoch} step {step + 1} global_step {global_step + 1} to {exp_dir}" + ) + + # the continue epochs are not resumed, so we need to reset the sampler start index and start step + dataloader.sampler.set_start_index(0) + start_step = 0 + + +if __name__ == "__main__": + main() diff --git a/scripts/train.sh b/scripts/train.sh deleted file mode 100644 index b38de612..00000000 --- a/scripts/train.sh +++ /dev/null @@ -1,44 +0,0 @@ -#!/usr/bin/env bash - -# get args -GPUS=${1:-8} - -# get root dir -FOLDER_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )" -ROOT_DIR=$FOLDER_DIR/.. - -# go to root dir -cd $ROOT_DIR - -# define dataset shards -COLLATED_VIDEO_DIR=./dataset/MSRVTT-collated/train/videos -PROCESSED_DATASET=( - ./dataset/MSRVTT-processed/train/part-00000 - ./dataset/MSRVTT-processed/train/part-00001 - ./dataset/MSRVTT-processed/train/part-00002 - ./dataset/MSRVTT-processed/train/part-00003 - ./dataset/MSRVTT-processed/train/part-00004 - ./dataset/MSRVTT-processed/train/part-00005 - ./dataset/MSRVTT-processed/train/part-00006 - ./dataset/MSRVTT-processed/train/part-00007 - ./dataset/MSRVTT-processed/train/part-00008 - ./dataset/MSRVTT-processed/train/part-00009 -) - -# create timestamp to differentiate between runs -timestamp=$(date +%Y-%m-%d-%H-%M) - -# run single node training -torchrun --standalone \ - --nproc_per_node $GPUS \ - train.py \ - --epochs 1 \ - --batch_size 1 \ - --lr 1e-4 \ - --accumulation_steps 32 \ - --grad_checkpoint \ - --dataset "${PROCESSED_DATASET[@]}" \ - --video_dir $COLLATED_VIDEO_DIR \ - --save_interval 224 \ - --checkpoint_dir ./checkpoints/$timestamp \ - --tensorboard_dir ./runs/$timestamp diff --git a/setup.py b/setup.py index 582beefb..45049bba 100644 --- a/setup.py +++ b/setup.py @@ -1,4 +1,3 @@ -import os from typing import List from setuptools import find_packages, setup @@ -29,47 +28,26 @@ def fetch_readme() -> str: return f.read() -def get_version() -> str: - """ - This function reads the version.txt and generates the colossalai/version.py file. - - Returns: - The library version stored in version.txt. - """ - - setup_file_path = os.path.abspath(__file__) - project_path = os.path.dirname(setup_file_path) - version_txt_path = os.path.join(project_path, "version.txt") - - with open(version_txt_path) as f: - version = f.read().strip() - return version - - setup( - name="open-sora", - version=get_version(), + name="opensora", + version="1.0.0", packages=find_packages( exclude=( - "docker", - "tests", + "assets", + "configs", "docs", - "examples", - "tests", + "outputs", + "pretrained_models", "scripts", - "requirements", - "extensions", + "tests", + "tools", "*.egg-info", - ), + ) ), - description="Unofficial implementation of OpenAI's Sora by the Colossal-AI Team", + description="Democratizing Efficient Video Production for All", long_description=fetch_readme(), long_description_content_type="text/markdown", license="Apache Software License 2.0", - url="https://www.colossalai.org", - project_urls={ - "Github": "https://github.com/hpcaitech/Open-Sora", - }, install_requires=fetch_requirements("requirements.txt"), python_requires=">=3.6", classifiers=[ diff --git a/tests/test_model.py b/tests/test_model.py deleted file mode 100644 index 1c2462fc..00000000 --- a/tests/test_model.py +++ /dev/null @@ -1,117 +0,0 @@ -import colossalai -import pytest -import torch -import torch.distributed as dist -import torch.nn.functional as F -from colossalai.booster import Booster -from colossalai.logging import disable_existing_loggers -from colossalai.nn.optimizer import HybridAdam -from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn -from colossalai.utils import get_current_device -from torch.testing import assert_close - -from open_sora.modeling import DiT_models -from open_sora.utils.plugin import ZeroSeqParallelPlugin - - -@parameterize("sp_size", [2, 4]) -@parameterize("sp_mode", ["ulysses", "fastseq"]) -@parameterize("sp_overlap", [True, False]) -@parameterize("cross_attn", [True, False]) -def check_dit_model_fwd_bwd( - sp_size: int, - sp_mode: str, - sp_overlap: bool, - cross_attn: bool, - video_latent_states, - text_latent_states, - t, - mask, -): - if sp_overlap and sp_mode == "ulysses": - return - if not cross_attn: - mask = mask.new_ones( - *mask.shape[:2], - mask.shape[2] + mask.shape[3], - mask.shape[2] + mask.shape[3], - ) - plugin = ZeroSeqParallelPlugin( - sp_size=sp_size, stage=2, precision="fp32", master_weights=False - ) - booster = Booster(plugin=plugin) - model = DiT_models["DiT-B/8"](text_dropout_prob=0.0, use_cross_attn=cross_attn).to( - get_current_device() - ) - parallel_model = DiT_models["DiT-B/8"]( - text_dropout_prob=0.0, - use_cross_attn=cross_attn, - seq_parallel_group=plugin.sp_group, - seq_parallel_mode=sp_mode, - seq_parallel_overlap=sp_overlap, - ).to(get_current_device()) - parallel_model.load_state_dict(model.state_dict()) - opt = HybridAdam(parallel_model.parameters(), lr=1e-3) - parallel_model, opt, *_ = booster.boost(parallel_model, opt) - - target = model(video_latent_states, t, text_latent_states, mask) - noise = torch.randn_like(target) - target_loss = F.mse_loss(target, noise) - target_loss.backward() - - dp_video_latent_states = video_latent_states.chunk(plugin.dp_size)[plugin.dp_rank] - dp_text_latent_states = text_latent_states.chunk(plugin.dp_size)[plugin.dp_rank] - dp_t = t.chunk(plugin.dp_size)[plugin.dp_rank] - dp_mask = mask.chunk(plugin.dp_size)[plugin.dp_rank] - dp_noise = noise.chunk(plugin.dp_size)[plugin.dp_rank] - - output = parallel_model( - dp_video_latent_states, dp_t, dp_text_latent_states, dp_mask - ) - loss = F.mse_loss(output, dp_noise) - booster.backward(loss, opt) - - if plugin.dp_size == 1: - assert_close(target, output) - - for p1, p2 in zip(model.parameters(), opt._master_param_groups_of_current_rank[0]): - working_p = opt._param_store.master_to_working_param[id(p2)] - grads = opt._grad_store.get_partitioned_gradients_by_param_id(0, id(working_p)) - grad_index = 0 if opt._partition_grads else opt._local_rank - grad = grads[grad_index] - sharded_grad = p1.grad.view(-1).chunk(dist.get_world_size())[dist.get_rank()] - assert_close(sharded_grad, grad[: sharded_grad.shape[0]]) - - -def run_dist(rank, world_size, port): - disable_existing_loggers() - colossalai.launch( - config={}, - rank=rank, - world_size=world_size, - port=port, - host="localhost", - backend="nccl", - ) - b, s, c, p = 4, 20, 3, 8 - dim_text, s_text = 512, 12 - video_latent_states = torch.rand(b, s, c, p, p, device=get_current_device()) - text_latent_states = torch.rand(b, s_text, dim_text, device=get_current_device()) - t = torch.randint(0, 1000, (b,), device=get_current_device()) - mask = torch.ones(b, 1, s, s_text, device=get_current_device(), dtype=torch.int) - check_dit_model_fwd_bwd( - video_latent_states=video_latent_states, - text_latent_states=text_latent_states, - t=t, - mask=mask, - ) - - -@pytest.mark.dist -@rerun_if_address_is_in_use() -def test_dit_model(): - spawn(run_dist, 4) - - -if __name__ == "__main__": - test_dit_model() diff --git a/tests/test_seq_parallel_attention.py b/tests/test_seq_parallel_attention.py new file mode 100644 index 00000000..00966ad0 --- /dev/null +++ b/tests/test_seq_parallel_attention.py @@ -0,0 +1,149 @@ +import colossalai +import torch +import torch.distributed as dist +from colossalai.testing import spawn + +from opensora.acceleration.communications import gather_forward_split_backward, split_forward_gather_backward +from opensora.acceleration.parallel_states import set_sequence_parallel_group +from opensora.models.layers.blocks import ( + Attention, + MultiHeadCrossAttention, + SeqParallelAttention, + SeqParallelMultiHeadCrossAttention, +) + + +def run_attention(rank, world_size): + # create model + torch.manual_seed(1024) + set_sequence_parallel_group(dist.group.WORLD) + + seq_parallel_attention = SeqParallelAttention(dim=256, num_heads=4, qkv_bias=True, enable_flashattn=False).cuda() + + torch.manual_seed(1024) + attention = Attention( + dim=256, + num_heads=4, + qkv_bias=True, + enable_flashattn=False, + ).cuda() + + # create inputs + torch.manual_seed(1024) + x = torch.randn(4, 64, 256).cuda() + seq_x = x.clone().detach() + + x.requires_grad = True + x.retain_grad() + seq_x.requires_grad = True + seq_x.retain_grad() + + sub_seq_x = split_forward_gather_backward(seq_x, dist.group.WORLD, dim=1, grad_scale="down") + + # run model + out = attention(x) + sub_seq_out = seq_parallel_attention(sub_seq_x) + seq_out = gather_forward_split_backward(sub_seq_out, dist.group.WORLD, dim=1, grad_scale="up") + + assert torch.allclose(seq_out, out, atol=1e-7), f"{seq_out}\nvs\n{out}" + + # run backward + seq_out.mean().backward() + out.mean().backward() + + # all reduce gradient for sp + for p in seq_parallel_attention.parameters(): + if p.grad is not None: + dist.all_reduce(p.grad, group=dist.group.WORLD) + p.grad.div_(world_size) + + # check grad + for p1, p2 in zip(seq_parallel_attention.parameters(), attention.parameters()): + assert torch.allclose(p1.grad, p2.grad, atol=1e-7), f"{p1.grad}\nvs\n{p2.grad}" + + # check input grad + assert torch.allclose(x.grad, seq_x.grad, atol=1e-7), f"{x.grad}\nvs\n{seq_x.grad}" + + +def run_cross_attention(rank, world_size): + # create model + torch.manual_seed(1024) + set_sequence_parallel_group(dist.group.WORLD) + seq_parallel_attention = SeqParallelMultiHeadCrossAttention( + d_model=256, + num_heads=4, + ).cuda().to(torch.bfloat16) + + torch.manual_seed(1024) + attention = MultiHeadCrossAttention( + d_model=256, + num_heads=4, + ).cuda().to(torch.bfloat16) + + # make sure the weights are the same + for p1, p2 in zip(seq_parallel_attention.parameters(), attention.parameters()): + p1.data.copy_(p2.data) + + # create inputs + torch.manual_seed(1024) + x = torch.randn(4, 64, 256).cuda().to(torch.bfloat16) + y = torch.randn(4, 32, 256).cuda().to(torch.bfloat16) + + mask = [2, 10, 8, 16] + mask = None + seq_x = x.clone().detach() + seq_y = y.clone().detach() + + # set grad + x.requires_grad = True + x.retain_grad() + seq_x.requires_grad = True + seq_x.retain_grad() + y.requires_grad = True + y.retain_grad() + seq_y.requires_grad = True + seq_y.retain_grad() + + # split by sequence + sub_seq_x = split_forward_gather_backward(seq_x, dist.group.WORLD, dim=1, grad_scale="down") + + # run model + out = attention(x, y, mask) + sub_seq_out = seq_parallel_attention(sub_seq_x, seq_y, mask) + seq_out = gather_forward_split_backward(sub_seq_out, dist.group.WORLD, dim=1, grad_scale="up") + + assert torch.allclose(seq_out, out, rtol=1e-5, atol=1e-6), f"\n{seq_out}\nvs\n{out}" + + # run backward + seq_out.mean().backward() + out.mean().backward() + + # all reduce gradient for sp + for name, p in seq_parallel_attention.named_parameters(): + if p.grad is not None: + dist.all_reduce(p.grad, group=dist.group.WORLD) + p.grad.div_(world_size) + else: + print(f"grad of {name} is None") + + # # check grad + for p1, p2 in zip(seq_parallel_attention.named_parameters(), attention.named_parameters()): + assert torch.allclose(p1[1].grad, p2[1].grad, rtol=1e-3, atol=1e-4), f"\n{p1[0]}\nvs\n{p2[0]}:\n{p1[1].grad}\nvs\n{p2[1].grad}" + + # # check input grad + assert torch.allclose(x.grad, seq_x.grad, atol=1e-7), f"{x.grad}\nvs\n{seq_x.grad}" + assert torch.allclose(y.grad, seq_y.grad, atol=1e-7), f"{y.grad}\nvs\n{seq_y.grad}" + + +def run_dist(rank, world_size, port): + colossalai.launch({}, rank=rank, world_size=world_size, host="localhost", port=port) + # run_attention(rank, world_size) + run_cross_attention(rank, world_size) + + +def test_seq_parallel_attention(): + spawn(run_dist, nprocs=2) + + +if __name__ == "__main__": + test_seq_parallel_attention() diff --git a/tests/test_sp_attn.py b/tests/test_sp_attn.py deleted file mode 100644 index d62a4ab5..00000000 --- a/tests/test_sp_attn.py +++ /dev/null @@ -1,102 +0,0 @@ -import colossalai -import pytest -import torch -import torch.distributed as dist -from colossalai.logging import disable_existing_loggers -from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn -from colossalai.utils import get_current_device -from torch.testing import assert_close - -from open_sora.modeling.dit import ( - CrossAttention, - FastSeqParallelCrossAttention, - SeqParallelCrossAttention, -) -from open_sora.utils.comm import gather_seq, split_seq - - -@parameterize("layer_cls", [SeqParallelCrossAttention, FastSeqParallelCrossAttention]) -@parameterize("overlap", [True, False]) -@parameterize("cross_attn", [True, False]) -def check_sp_attn(layer_cls, overlap, cross_attn): - if overlap and layer_cls == SeqParallelCrossAttention: - return - model_kwargs = {} - if layer_cls == FastSeqParallelCrossAttention: - model_kwargs["overlap"] = overlap - sp_size = dist.get_world_size() - sp_rank = dist.get_rank() - q_dim, context_dim = 8, 4 - num_heads = 4 - head_dim = 16 - bs = 2 - sq = 8 - skv = 4 - if not cross_attn: - context_dim = None - skv = sq - attn = CrossAttention(q_dim, context_dim, num_heads, head_dim).to( - get_current_device() - ) - parallel_attn = layer_cls( - q_dim, - context_dim, - num_heads, - head_dim, - seq_parallel_group=dist.group.WORLD, - **model_kwargs, - ).to(get_current_device()) - parallel_attn.load_state_dict(attn.state_dict()) - hidden_states = torch.rand(bs, sq, q_dim, device=get_current_device()) - if cross_attn: - context = torch.rand(bs, skv, context_dim, device=get_current_device()) - else: - context = None - mask = torch.zeros(bs, 1, sq, skv, device=get_current_device()) - target = attn(hidden_states, context, mask) - hidden_states_parallel = split_seq(hidden_states, sp_size, sp_rank) - if cross_attn: - context_parallel = split_seq(context, sp_size, sp_rank) - else: - context_parallel = None - output_parallel = parallel_attn( - hidden_states_parallel, - context_parallel, - mask, - ) - assert torch.equal(target.chunk(sp_size, dim=1)[sp_rank], output_parallel) - output = gather_seq(output_parallel, sp_size, sp_rank, dist.group.WORLD) - assert torch.equal(target, output) - target.mean().backward() - output.mean().backward() - - # all-reduce mean of grads - for p in parallel_attn.parameters(): - p.grad.data.div_(sp_size) - dist.all_reduce(p.grad.data) - - for (n, p1), p2 in zip(attn.named_parameters(), parallel_attn.parameters()): - assert_close(p1.grad, p2.grad, msg=lambda m: f"Check {n}\n{m}") - - -def run_dist(rank, world_size, port): - disable_existing_loggers() - colossalai.launch( - config={}, - rank=rank, - world_size=world_size, - port=port, - host="localhost", - backend="nccl", - ) - check_sp_attn() - - -@pytest.mark.dist -@rerun_if_address_is_in_use() -def test_seq_parallel_attn(): - spawn(run_dist, 2) - - -if __name__ == "__main__": - test_seq_parallel_attn() diff --git a/tests/test_t5_shardformer.py b/tests/test_t5_shardformer.py new file mode 100644 index 00000000..68040ab3 --- /dev/null +++ b/tests/test_t5_shardformer.py @@ -0,0 +1,71 @@ +import time +from copy import deepcopy + +import colossalai +import torch +from colossalai.shardformer import ShardConfig, ShardFormer +from colossalai.testing import spawn + +from opensora.acceleration.shardformer.policy.t5_encoder import T5EncoderPolicy +from opensora.models.text_encoder.t5 import T5Embedder + + +def run_t5_encoder(rank, world_size, port): + colossalai.launch({}, rank=rank, world_size=world_size, port=port, host="localhost") + + # t5 embedder + t5_path = "./pretrained_models/t5_ckpts" + hf_t5 = T5Embedder(device="cuda", local_cache=True, cache_dir=t5_path, torch_dtype=torch.float) + sf_t5 = deepcopy(hf_t5) + + # create huggingface model as normal + shard_config = ShardConfig( + tensor_parallel_process_group=None, + pipeline_stage_manager=None, + enable_tensor_parallelism=False, + enable_fused_normalization=False, + enable_flash_attention=False, + enable_jit_fused=True, + enable_sequence_parallelism=False, + enable_sequence_overlap=False, + ) + shard_former = ShardFormer(shard_config=shard_config) + sharded_model, _ = shard_former.optimize(sf_t5.model, policy=T5EncoderPolicy()) + sf_t5.model = sharded_model + + # test t5 embedder + texts = ["Who is the best player in the history of NBA?", "How to study computer science?"] + for i in range(5): + hf_embs, hf_masks = hf_t5.get_text_embeddings(texts) + sf_embs, sf_masks = sf_t5.get_text_embeddings(texts) + + # check accuracy + assert torch.allclose(hf_embs, sf_embs, rtol=1e-4, atol=1e-5), f"{hf_embs} \nvs\n{sf_embs}" + assert torch.allclose(hf_masks, sf_masks), f"{hf_masks} \nvs\n{sf_masks}" + + # measure perf + torch.cuda.synchronize() + hf_start = time.time() + for i in range(20): + hf_embs, hf_masks = hf_t5.get_text_embeddings(texts) + torch.cuda.synchronize() + hf_end = time.time() + + # convert sf to fp16 + hf_t5.model = hf_t5.model.half() + torch.cuda.synchronize() + sf_start = time.time() + for i in range(20): + hf_embs, hf_masks = hf_t5.get_text_embeddings(texts) + torch.cuda.synchronize() + sf_end = time.time() + + print(f"[Performance] native: {hf_end - hf_start}s, shardformer: {sf_end - sf_start} s") + + +def test_t5_encoder(): + spawn(run_t5_encoder) + + +if __name__ == "__main__": + test_t5_encoder() diff --git a/tools/__init__.py b/tools/__init__.py new file mode 100644 index 00000000..e69de29b diff --git a/tools/caption/README.md b/tools/caption/README.md new file mode 100644 index 00000000..e9289d05 --- /dev/null +++ b/tools/caption/README.md @@ -0,0 +1,25 @@ +# Video Captioning + +Human labeling of videos is expensive and time-consuming. We adopt powerful image captioning models to generate captions for videos. Although GPT-4V achieves a better performance, its 20s/sample speed is too slow for us. With batch inference, we can achieve a speed of 3s/sample with LLaVA, and the quality is comparable. LLaVA is the second best open-source model in [MMMU](https://mmmu-benchmark.github.io/) and accepts any resolution. + +![Caption](https://i0.imgs.ovh/2024/03/16/eXdvC.png) + +## GPT-4V Captioning + +Run the following command to generate captions for videos with GPT-4V: + +```bash +python -m tools.caption.caption_gpt4 FOLDER_WITH_VIDEOS output.csv --key $OPENAI_API_KEY +``` + +The cost is approximately $0.01 per video (3 frames per video). The output is a CSV file with path and caption. + +## LLaVA Captioning + +First, install LLaVA according to their [official instructions](https://github.com/haotian-liu/LLaVA?tab=readme-ov-file#install). We use the `liuhaotian/llava-v1.6-34b` model for captioning, which can be download [here](https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b). Then, run the following command to generate captions for videos with LLaVA: + +```bash +CUDA_VISIBLE_DEVICES=0,1 python -m tools.caption.caption_llava samples output.csv +``` + +The Yi-34B requires 2 80GB GPUs and 3s/sample. The output is a CSV file with path and caption. diff --git a/tools/caption/__init__.py b/tools/caption/__init__.py new file mode 100644 index 00000000..e69de29b diff --git a/tools/caption/caption_gpt4.py b/tools/caption/caption_gpt4.py new file mode 100644 index 00000000..b2c7590f --- /dev/null +++ b/tools/caption/caption_gpt4.py @@ -0,0 +1,69 @@ +import argparse +import csv +import os + +import requests +import tqdm + +from .utils import extract_frames, prompts, read_video_list + + +def get_caption(frame, prompt, api_key): + headers = {"Content-Type": "application/json", "Authorization": f"Bearer {api_key}"} + payload = { + "model": "gpt-4-vision-preview", + "messages": [ + { + "role": "user", + "content": [ + { + "type": "text", + "text": prompt, + }, + {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{frame[0]}"}}, + {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{frame[1]}"}}, + {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{frame[2]}"}}, + ], + } + ], + "max_tokens": 300, + } + response = requests.post("https://api.openai.com/v1/chat/completions", headers=headers, json=payload, timeout=60) + caption = response.json()["choices"][0]["message"]["content"] + caption = caption.replace("\n", " ") + return caption + + +def main(args): + # ====================================================== + # 1. read video list + # ====================================================== + videos = read_video_list(args.video_folder, args.output_file) + f = open(args.output_file, "a") + writer = csv.writer(f) + + # ====================================================== + # 2. generate captions + # ====================================================== + for video in tqdm.tqdm(videos): + video_path = os.path.join(args.video_folder, video) + frame, length = extract_frames(video_path, base_64=True) + if len(frame) < 3: + continue + + prompt = prompts[args.prompt] + caption = get_caption(frame, prompt, args.key) + + writer.writerow((video, caption, length)) + f.close() + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument("video_folder", type=str) + parser.add_argument("output_file", type=str) + parser.add_argument("--prompt", type=str, default="three_frames") + parser.add_argument("--key", type=str) + args = parser.parse_args() + + main(args) diff --git a/tools/caption/caption_llava.py b/tools/caption/caption_llava.py new file mode 100644 index 00000000..8f4278c3 --- /dev/null +++ b/tools/caption/caption_llava.py @@ -0,0 +1,352 @@ +import argparse +import csv +import os +import warnings + +import torch +from llava.constants import DEFAULT_IMAGE_TOKEN, IGNORE_INDEX, IMAGE_TOKEN_INDEX +from llava.conversation import conv_templates +from llava.mm_utils import get_anyres_image_grid_shape, get_model_name_from_path, process_images, tokenizer_image_token +from llava.model.builder import load_pretrained_model +from llava.model.llava_arch import unpad_image +from llava.utils import disable_torch_init +from tqdm import tqdm + +from .utils import extract_frames, prompts, read_video_list + +disable_torch_init() + + +def prepare_inputs_labels_for_multimodal( + self, input_ids, position_ids, attention_mask, past_key_values, labels, images, image_sizes=None +): + # llava_arch.py + vision_tower = self.get_vision_tower() + if vision_tower is None or images is None or input_ids.shape[1] == 1: + return input_ids, position_ids, attention_mask, past_key_values, None, labels + + if type(images) is list or images.ndim == 5: + if type(images) is list: + images = [x.unsqueeze(0) if x.ndim == 3 else x for x in images] + concat_images = torch.cat([image for image in images], dim=0) + image_features = self.encode_images(concat_images) + split_sizes = [image.shape[0] for image in images] + image_features = torch.split(image_features, split_sizes, dim=0) + mm_patch_merge_type = getattr(self.config, "mm_patch_merge_type", "flat") + image_aspect_ratio = getattr(self.config, "image_aspect_ratio", "square") + if mm_patch_merge_type == "flat": + image_features = [x.flatten(0, 1) for x in image_features] + elif mm_patch_merge_type.startswith("spatial"): + new_image_features = [] + for image_idx, image_feature in enumerate(image_features): + if image_feature.shape[0] > 1: + base_image_feature = image_feature[0] + image_feature = image_feature[1:] + height = width = self.get_vision_tower().num_patches_per_side + assert height * width == base_image_feature.shape[0] + if image_aspect_ratio == "anyres": + num_patch_width, num_patch_height = get_anyres_image_grid_shape( + image_sizes[image_idx], + self.config.image_grid_pinpoints, + self.get_vision_tower().config.image_size, + ) + image_feature = image_feature.view(num_patch_height, num_patch_width, height, width, -1) + else: + raise NotImplementedError + if "unpad" in mm_patch_merge_type: + image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous() + image_feature = image_feature.flatten(1, 2).flatten(2, 3) + image_feature = unpad_image(image_feature, image_sizes[image_idx]) + image_feature = torch.cat( + ( + image_feature, + self.model.image_newline[:, None, None] + .expand(*image_feature.shape[:-1], 1) + .to(image_feature.device), + ), + dim=-1, + ) + image_feature = image_feature.flatten(1, 2).transpose(0, 1) + else: + image_feature = image_feature.permute(0, 2, 1, 3, 4).contiguous() + image_feature = image_feature.flatten(0, 3) + image_feature = torch.cat((base_image_feature, image_feature), dim=0) + else: + image_feature = image_feature[0] + if "unpad" in mm_patch_merge_type: + image_feature = torch.cat( + (image_feature, self.model.image_newline[None].to(image_feature.device)), dim=0 + ) + new_image_features.append(image_feature) + image_features = new_image_features + else: + raise ValueError(f"Unexpected mm_patch_merge_type: {self.config.mm_patch_merge_type}") + else: + image_features = self.encode_images(images) + + # TODO: image start / end is not implemented here to support pretraining. + if getattr(self.config, "tune_mm_mlp_adapter", False) and getattr(self.config, "mm_use_im_start_end", False): + raise NotImplementedError + + # Let's just add dummy tensors if they do not exist, + # it is a headache to deal with None all the time. + # But it is not ideal, and if you have a better idea, + # please open an issue / submit a PR, thanks. + _labels = labels + _position_ids = position_ids + _attention_mask = attention_mask + if attention_mask is None: + attention_mask = torch.ones_like(input_ids, dtype=torch.bool) + else: + attention_mask = attention_mask.bool() + if position_ids is None: + position_ids = torch.arange(0, input_ids.shape[1], dtype=torch.long, device=input_ids.device) + if labels is None: + labels = torch.full_like(input_ids, IGNORE_INDEX) + + # remove the padding using attention_mask -- FIXME + input_ids = [ + cur_input_ids[cur_attention_mask] for cur_input_ids, cur_attention_mask in zip(input_ids, attention_mask) + ] + labels = [cur_labels[cur_attention_mask] for cur_labels, cur_attention_mask in zip(labels, attention_mask)] + + new_input_embeds = [] + new_labels = [] + cur_image_idx = 0 + for batch_idx, cur_input_ids in enumerate(input_ids): + num_images = (cur_input_ids == IMAGE_TOKEN_INDEX).sum() + if num_images == 0: + cur_image_features = image_features[cur_image_idx] + cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids) + cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0]], dim=0) + new_input_embeds.append(cur_input_embeds) + new_labels.append(labels[batch_idx]) + cur_image_idx += 1 + continue + + image_token_indices = ( + [-1] + torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0].tolist() + [cur_input_ids.shape[0]] + ) + cur_input_ids_noim = [] + cur_labels = labels[batch_idx] + cur_labels_noim = [] + for i in range(len(image_token_indices) - 1): + cur_input_ids_noim.append(cur_input_ids[image_token_indices[i] + 1 : image_token_indices[i + 1]]) + cur_labels_noim.append(cur_labels[image_token_indices[i] + 1 : image_token_indices[i + 1]]) + split_sizes = [x.shape[0] for x in cur_labels_noim] + cur_input_embeds = self.get_model().embed_tokens(torch.cat(cur_input_ids_noim)) + cur_input_embeds_no_im = torch.split(cur_input_embeds, split_sizes, dim=0) + cur_new_input_embeds = [] + cur_new_labels = [] + + for i in range(num_images + 1): + cur_new_input_embeds.append(cur_input_embeds_no_im[i]) + cur_new_labels.append(cur_labels_noim[i]) + if i < num_images: + cur_image_features = image_features[cur_image_idx] + cur_image_idx += 1 + cur_new_input_embeds.append(cur_image_features) + cur_new_labels.append( + torch.full( + (cur_image_features.shape[0],), + IGNORE_INDEX, + device=cur_labels.device, + dtype=cur_labels.dtype, + ) + ) + + cur_new_input_embeds = [x.to(self.device) for x in cur_new_input_embeds] + + cur_new_input_embeds = torch.cat(cur_new_input_embeds) + cur_new_labels = torch.cat(cur_new_labels) + + new_input_embeds.append(cur_new_input_embeds) + new_labels.append(cur_new_labels) + + # Truncate sequences to max length as image embeddings can make the sequence longer + tokenizer_model_max_length = getattr(self.config, "tokenizer_model_max_length", None) + if tokenizer_model_max_length is not None: + new_input_embeds = [x[:tokenizer_model_max_length] for x in new_input_embeds] + new_labels = [x[:tokenizer_model_max_length] for x in new_labels] + + # Combine them + max_len = max(x.shape[0] for x in new_input_embeds) + batch_size = len(new_input_embeds) + + new_input_embeds_padded = [] + new_labels_padded = torch.full( + (batch_size, max_len), IGNORE_INDEX, dtype=new_labels[0].dtype, device=new_labels[0].device + ) + attention_mask = torch.zeros((batch_size, max_len), dtype=attention_mask.dtype, device=attention_mask.device) + position_ids = torch.zeros((batch_size, max_len), dtype=position_ids.dtype, device=position_ids.device) + + for i, (cur_new_embed, cur_new_labels) in enumerate(zip(new_input_embeds, new_labels)): + cur_len = cur_new_embed.shape[0] + if getattr(self.config, "tokenizer_padding_side", "right") == "left": + new_input_embeds_padded.append( + torch.cat( + ( + torch.zeros( + (max_len - cur_len, cur_new_embed.shape[1]), + dtype=cur_new_embed.dtype, + device=cur_new_embed.device, + ), + cur_new_embed, + ), + dim=0, + ) + ) + if cur_len > 0: + new_labels_padded[i, -cur_len:] = cur_new_labels + attention_mask[i, -cur_len:] = True + position_ids[i, -cur_len:] = torch.arange( + 0, cur_len, dtype=position_ids.dtype, device=position_ids.device + ) + else: + new_input_embeds_padded.append( + torch.cat( + ( + cur_new_embed, + torch.zeros( + (max_len - cur_len, cur_new_embed.shape[1]), + dtype=cur_new_embed.dtype, + device=cur_new_embed.device, + ), + ), + dim=0, + ) + ) + if cur_len > 0: + new_labels_padded[i, :cur_len] = cur_new_labels + attention_mask[i, :cur_len] = True + position_ids[i, :cur_len] = torch.arange( + 0, cur_len, dtype=position_ids.dtype, device=position_ids.device + ) + + new_input_embeds = torch.stack(new_input_embeds_padded, dim=0) + + if _labels is None: + new_labels = None + else: + new_labels = new_labels_padded + + if _attention_mask is None: + attention_mask = None + else: + attention_mask = attention_mask.to(dtype=_attention_mask.dtype) + + if _position_ids is None: + position_ids = None + + return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels + + +@torch.inference_mode() +def main(args): + # ====================================================== + # 1. read video list + # ====================================================== + videos = read_video_list(args.video_folder, args.output_file) + f = open(args.output_file, "a") + writer = csv.writer(f) + + # ====================================================== + # 2. load model and prepare prompts + # ====================================================== + model_path = "liuhaotian/llava-v1.6-34b" + query = prompts[args.prompt] + print(f"Prompt: {query}") + conv = conv_templates["chatml_direct"].copy() + conv.append_message(conv.roles[0], DEFAULT_IMAGE_TOKEN + "\n" + query) + prompt = conv.get_prompt() + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") # Pytorch non-meta copying warning fills out the console + tokenizer, model, image_processor, context_len = load_pretrained_model( + model_path=model_path, + model_base=None, + model_name=get_model_name_from_path(model_path), + ) + input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt") + input_ids = input_ids.unsqueeze(0).to(model.device) + + # ====================================================== + # 3. generate captions + # ====================================================== + bs = args.bs + for i in tqdm(range(0, len(videos), bs)): + # prepare a batch of inputs + video_files = videos[i : i + bs] + frames = [] + video_lengths = [] + for video_file in video_files: + frame, length = extract_frames(os.path.join(args.video_folder, video_file)) + if len(frame) < 3: + continue + frames.append(frame) + video_lengths.append(length) + if len(frames) == 0: + continue + + # encode the batch of inputs + samples = [] + for imgs in frames: + imgs_size = [img.size for img in imgs] + imgs = process_images(imgs, image_processor, model.config) + imgs = imgs.to(model.device, dtype=torch.float16) + with torch.inference_mode(): + _, _, _, _, inputs_embeds, _ = prepare_inputs_labels_for_multimodal( + model, input_ids, None, None, None, None, images=imgs, image_sizes=imgs_size + ) + samples.append(inputs_embeds) + + # padding + max_len = max([sample.shape[1] for sample in samples]) + attention_mask = torch.tensor( + [[0] * (max_len - samples[i].shape[1]) + [1] * samples[i].shape[1] for i in range(len(samples))] + ).to(model.device) + inputs_embeds = [ + torch.cat( + [ + torch.zeros( + (1, max_len - samples[i].shape[1], samples[i].shape[-1]), + device=model.device, + dtype=torch.float16, + ), + samples[i], + ], + dim=1, + ) + for i in range(len(samples)) + ] + inputs_embeds = torch.cat(inputs_embeds, dim=0) + + # generate outputs + output_ids = super(type(model), model).generate( + inputs_embeds=inputs_embeds, + attention_mask=attention_mask, + do_sample=True, + temperature=0.2, + max_new_tokens=512, + use_cache=True, + ) + outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True) + outputs = [output.replace("\n", " ").strip() for output in outputs] + + # save results + result = list(zip(video_files, outputs, video_lengths)) + for t in result: + writer.writerow(t) + + f.close() + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument("video_folder", type=str) + parser.add_argument("output_file", type=str) + parser.add_argument("--bs", type=int, default=32) + parser.add_argument("--prompt", type=str, default="three_frames") + args = parser.parse_args() + + main(args) diff --git a/tools/caption/utils.py b/tools/caption/utils.py new file mode 100644 index 00000000..3912f0cc --- /dev/null +++ b/tools/caption/utils.py @@ -0,0 +1,67 @@ +import base64 +import csv +import os + +import cv2 +from PIL import Image + +prompts = { + "naive": "Describe the video", + "three_frames": "A video is given by providing three frames in chronological order. Describe this video and its style to generate a description. Pay attention to all objects in the video. Do not describe each frame individually. Do not reply with words like 'first frame'. The description should be useful for AI to re-generate the video. The description should be less than six sentences. Here are some examples of good descriptions: 1. A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. She wears a black leather jacket, a long red dress, and black boots, and carries a black purse. She wears sunglasses and red lipstick. She walks confidently and casually. The street is damp and reflective, creating a mirror effect of the colorful lights. Many pedestrians walk about. 2. Several giant wooly mammoths approach treading through a snowy meadow, their long wooly fur lightly blows in the wind as they walk, snow covered trees and dramatic snow capped mountains in the distance, mid afternoon light with wispy clouds and a sun high in the distance creates a warm glow, the low camera view is stunning capturing the large furry mammal with beautiful photography, depth of field. 3. Drone view of waves crashing against the rugged cliffs along Big Sur's garay point beach. The crashing blue waters create white-tipped waves, while the golden light of the setting sun illuminates the rocky shore. A small island with a lighthouse sits in the distance, and green shrubbery covers the cliff's edge. The steep drop from the road down to the beach is a dramatic feat, with the cliff’s edges jutting out over the sea. This is a view that captures the raw beauty of the coast and the rugged landscape of the Pacific Coast Highway.", +} + + +def get_filelist(file_path): + Filelist = [] + VID_EXTENSIONS = ("mp4", "avi", "mov", "mkv") + for home, dirs, files in os.walk(file_path): + for filename in files: + ext = filename.split(".")[-1] + if ext in VID_EXTENSIONS: + Filelist.append(filename) + return Filelist + + +def get_video_length(cap): + return int(cap.get(cv2.CAP_PROP_FRAME_COUNT)) + + +def encode_image(image_path): + with open(image_path, "rb") as image_file: + return base64.b64encode(image_file.read()).decode("utf-8") + + +def extract_frames(video_path, points=(0.2, 0.5, 0.8), base_64=False): + cap = cv2.VideoCapture(video_path) + length = get_video_length(cap) + points = [int(length * point) for point in points] + frames = [] + if length < 3: + return frames, length + for point in points: + cap.set(cv2.CAP_PROP_POS_FRAMES, point) + ret, frame = cap.read() + if not base_64: + frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB) + frame = Image.fromarray(frame) + else: + _, buffer = cv2.imencode(".jpg", frame) + frame = base64.b64encode(buffer).decode("utf-8") + frames.append(frame) + return frames, length + + +def read_video_list(video_folder, output_file): + processed_videos = [] + if os.path.exists(output_file): + with open(output_file, "r") as f: + reader = csv.reader(f) + samples = list(reader) + processed_videos = [sample[0] for sample in samples] + + # read video list + videos = get_filelist(video_folder) + print(f"Dataset contains {len(videos)} videos.") + videos = [video for video in videos if video not in processed_videos] + print(f"Processing {len(videos)} new videos.") + return videos diff --git a/tools/datasets/README.md b/tools/datasets/README.md new file mode 100644 index 00000000..0118c8f1 --- /dev/null +++ b/tools/datasets/README.md @@ -0,0 +1,48 @@ +# Dataset Download and Management + +## Dataset Format + +The training data should be provided in a CSV file with the following format: + +```csv +/absolute/path/to/image1.jpg, caption1, num_of_frames +/absolute/path/to/image2.jpg, caption2, num_of_frames +``` + +## HD-VG-130M + +This dataset comprises 130M text-video pairs. You can download the dataset and prepare it for training according to [the dataset repository's instructions](https://github.com/daooshee/HD-VG-130M). There is a README.md file in the Google Drive link that provides instructions on how to download and cut the videos. For this version, we directly use the dataset provided by the authors. + +## Demo Dataset + +You can use ImageNet and UCF101 for a quick demo. After downloading the datasets, you can use the following command to prepare the csv file for the dataset: + +```bash +# ImageNet +python -m tools.datasets.convert_dataset imagenet IMAGENET_FOLDER --split train +# UCF101 +python -m tools.datasets.convert_dataset ucf101 UCF101_FOLDER --split videos +``` + +## Manage datasets + +We provide `csvutils.py` to manage the CSV files. You can use the following commands to process the CSV files: + +```bash +# generate DATA_fmin_128_fmax_256.csv with frames between 128 and 256 +python -m tools.datasets.csvutil DATA.csv --fmin 128 --fmax 256 +# generate DATA_root.csv with absolute path +python -m tools.datasets.csvutil DATA.csv --root /absolute/path/to/dataset +# remove videos with no captions +python -m tools.datasets.csvutil DATA.csv --remove-empty-caption +# compute the number of frames for each video +python -m tools.datasets.csvutil DATA.csv --relength +# remove caption prefix +python -m tools.datasets.csvutil DATA.csv --remove-caption-prefix +``` + +To merge multiple CSV files, you can use the following command: + +```bash +cat *csv > combined.csv +``` diff --git a/tools/datasets/__init__.py b/tools/datasets/__init__.py new file mode 100644 index 00000000..e69de29b diff --git a/tools/datasets/convert_dataset.py b/tools/datasets/convert_dataset.py new file mode 100644 index 00000000..4ff904fc --- /dev/null +++ b/tools/datasets/convert_dataset.py @@ -0,0 +1,66 @@ +import argparse +import csv +import os + +from torchvision.datasets import ImageNet + + +def get_filelist(file_path): + Filelist = [] + for home, dirs, files in os.walk(file_path): + for filename in files: + Filelist.append(os.path.join(home, filename)) + return Filelist + + +def split_by_capital(name): + # BoxingPunchingBag -> Boxing Punching Bag + new_name = "" + for i in range(len(name)): + if name[i].isupper() and i != 0: + new_name += " " + new_name += name[i] + return new_name + + +def process_imagenet(root, split): + root = os.path.expanduser(root) + data = ImageNet(root, split=split) + samples = [(path, data.classes[label][0]) for path, label in data.samples] + output = f"imagenet_{split}.csv" + + with open(output, "w") as f: + writer = csv.writer(f) + writer.writerows(samples) + + print(f"Saved {len(samples)} samples to {output}.") + + +def process_ucf101(root, split): + root = os.path.expanduser(root) + video_lists = get_filelist(os.path.join(root, split)) + classes = [x.split("/")[-2] for x in video_lists] + classes = [split_by_capital(x) for x in classes] + samples = list(zip(video_lists, classes)) + output = f"ucf101_{split}.csv" + + with open(output, "w") as f: + writer = csv.writer(f) + writer.writerows(samples) + + print(f"Saved {len(samples)} samples to {output}.") + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument("dataset", type=str, choices=["imagenet", "ucf101"]) + parser.add_argument("root", type=str) + parser.add_argument("--split", type=str, default="train") + args = parser.parse_args() + + if args.dataset == "imagenet": + process_imagenet(args.root, args.split) + elif args.dataset == "ucf101": + process_ucf101(args.root, args.split) + else: + raise ValueError("Invalid dataset") diff --git a/tools/datasets/csvutil.py b/tools/datasets/csvutil.py new file mode 100644 index 00000000..4bbd22db --- /dev/null +++ b/tools/datasets/csvutil.py @@ -0,0 +1,96 @@ +import argparse +import csv +import os + +from tqdm import tqdm + +# path, name, #frames +PREFIX = [ + "The video shows", + "The video captures", + "The video features", + "The video depicts", + "The video presents", + "The video features", + "The video is ", + "In the video,", +] + + +def get_video_length(path): + import cv2 + + cap = cv2.VideoCapture(path) + return int(cap.get(cv2.CAP_PROP_FRAME_COUNT)) + + +def main(args): + input_path = args.input + output_path = args.output + if output_path is None: + name = os.path.basename(input_path) + name, ext = os.path.splitext(name) + if args.fmin is not None: + name += f"_fmin_{args.fmin}" + if args.fmax is not None: + name += f"_fmax_{args.fmax}" + if args.remove_empty_caption: + name += "_rec" + if args.remove_caption_prefix: + name += "_rcp" + if args.root is not None: + name += f"_root" + if args.relength: + name += "_relength" + output_path = os.path.join(os.path.dirname(input_path), name + ext) + + with open(input_path, "r") as f: + reader = csv.reader(f) + data = list(reader) + print("Number of videos before filtering:", len(data)) + + data_new = [] + for i, row in tqdm(enumerate(data)): + path = row[0] + caption = row[1] + n_frames = int(row[2]) + if args.fmin is not None and n_frames < args.fmin: + continue + if args.fmax is not None and n_frames > args.fmax: + continue + if args.remove_empty_caption and len(caption) == 0: + continue + if args.remove_caption_prefix: + for prefix in PREFIX: + if caption.startswith(prefix): + caption = caption[len(prefix) :].strip() + if caption[0].islower(): + caption = caption[0].upper() + caption[1:] + row[1] = caption + break + if args.root is not None: + row[0] = os.path.join(args.root, path) + if args.relength: + n_frames = get_video_length(row[0]) + row[2] = n_frames + data_new.append(row) + + print("Number of videos after filtering:", len(data_new)) + with open(output_path, "w") as f: + writer = csv.writer(f) + writer.writerows(data_new) + print("Output saved to", output_path) + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument("input", type=str) + parser.add_argument("--output", type=str, default=None) + parser.add_argument("--fmin", type=int, default=None) + parser.add_argument("--fmax", type=int, default=None) + parser.add_argument("--root", type=str, default=None) + parser.add_argument("--remove-empty-caption", action="store_true") + parser.add_argument("--remove-caption-prefix", action="store_true") + parser.add_argument("--relength", action="store_true") + args = parser.parse_args() + main(args) diff --git a/tools/intepolate/README.md b/tools/intepolate/README.md new file mode 100644 index 00000000..e69de29b diff --git a/tools/scenedetect/README.md b/tools/scenedetect/README.md new file mode 100644 index 00000000..74658e57 --- /dev/null +++ b/tools/scenedetect/README.md @@ -0,0 +1,2 @@ +# Data Preprocessing Tools + diff --git a/tools/scenedetect/scene_detect.py b/tools/scenedetect/scene_detect.py new file mode 100644 index 00000000..c46e59d5 --- /dev/null +++ b/tools/scenedetect/scene_detect.py @@ -0,0 +1,138 @@ +import os +from multiprocessing import Pool + +from mmengine.logging import MMLogger +from scenedetect import ContentDetector, detect +from tqdm import tqdm + +from opensora.utils.misc import get_timestamp + +from .utils import check_mp4_integrity, clone_folder_structure, iterate_files, split_video + +# config +target_fps = 30 # int +shorter_size = 512 # int +min_seconds = 1 # float +max_seconds = 5 # float +assert max_seconds > min_seconds +cfg = dict( + target_fps=target_fps, + min_seconds=min_seconds, + max_seconds=max_seconds, + shorter_size=shorter_size, +) + + +def process_folder(root_src, root_dst): + # create logger + folder_path_log = os.path.dirname(root_dst) + log_name = os.path.basename(root_dst) + timestamp = get_timestamp() + log_path = os.path.join(folder_path_log, f"{log_name}_{timestamp}.log") + logger = MMLogger.get_instance(log_name, log_file=log_path) + + # clone folder structure + clone_folder_structure(root_src, root_dst) + + # all source videos + mp4_list = [x for x in iterate_files(root_src) if x.endswith(".mp4")] + mp4_list = sorted(mp4_list) + + for idx, sample_path in tqdm(enumerate(mp4_list)): + folder_src = os.path.dirname(sample_path) + folder_dst = os.path.join(root_dst, os.path.relpath(folder_src, root_src)) + + # check src video integrity + if not check_mp4_integrity(sample_path, logger=logger): + continue + + # detect scenes + scene_list = detect(sample_path, ContentDetector(), start_in_scene=True) + + # split scenes + save_path_list = split_video(sample_path, scene_list, save_dir=folder_dst, **cfg, logger=logger) + + # check integrity of generated clips + for x in save_path_list: + check_mp4_integrity(x, logger=logger) + + +def scene_detect(): + """detect & cut scenes using a single process + Expected dataset structure: + data/ + your_dataset/ + raw_videos/ + xxx.mp4 + yyy.mp4 + + This function results in: + data/ + your_dataset/ + raw_videos/ + xxx.mp4 + yyy.mp4 + zzz.mp4 + clips/ + xxx_scene-0.mp4 + yyy_scene-0.mp4 + yyy_scene-1.mp4 + """ + # TODO: specify your dataset root + root_src = f"./data/your_dataset/raw_videos" + root_dst = f"./data/your_dataset/clips" + + process_folder(root_src, root_dst) + + +def scene_detect_mp(): + """detect & cut scenes using multiple processes + Expected dataset structure: + data/ + your_dataset/ + raw_videos/ + split_0/ + xxx.mp4 + yyy.mp4 + split_1/ + xxx.mp4 + yyy.mp4 + + This function results in: + data/ + your_dataset/ + raw_videos/ + split_0/ + xxx.mp4 + yyy.mp4 + split_1/ + xxx.mp4 + yyy.mp4 + clips/ + split_0/ + xxx_scene-0.mp4 + yyy_scene-0.mp4 + split_1/ + xxx_scene-0.mp4 + yyy_scene-0.mp4 + yyy_scene-1.mp4 + """ + # TODO: specify your dataset root + root_src = f"./data/your_dataset/raw_videos" + root_dst = f"./data/your_dataset/clips" + + # TODO: specify your splits + splits = ["split_0", "split_1"] + + # process folders + root_src_list = [os.path.join(root_src, x) for x in splits] + root_dst_list = [os.path.join(root_dst, x) for x in splits] + + with Pool(processes=len(splits)) as pool: + pool.starmap(process_folder, list(zip(root_src_list, root_dst_list))) + + +if __name__ == "__main__": + # TODO: choose single process or multiprocessing + scene_detect() + # scene_detect_mp() diff --git a/tools/scenedetect/utils.py b/tools/scenedetect/utils.py new file mode 100644 index 00000000..19eae314 --- /dev/null +++ b/tools/scenedetect/utils.py @@ -0,0 +1,145 @@ +import os +import subprocess + +import cv2 +from imageio_ffmpeg import get_ffmpeg_exe +from mmengine.logging import print_log +from moviepy.editor import VideoFileClip +from scenedetect import FrameTimecode + + +def iterate_files(folder_path): + for root, dirs, files in os.walk(folder_path): + # root contains the current directory path + # dirs contains the list of subdirectories in the current directory + # files contains the list of files in the current directory + + # Process files in the current directory + for file in files: + file_path = os.path.join(root, file) + # print("File:", file_path) + yield file_path + + # Process subdirectories and recursively call the function + for subdir in dirs: + subdir_path = os.path.join(root, subdir) + # print("Subdirectory:", subdir_path) + iterate_files(subdir_path) + + +def iterate_folders(folder_path): + for root, dirs, files in os.walk(folder_path): + for subdir in dirs: + subdir_path = os.path.join(root, subdir) + yield subdir_path + # print("Subdirectory:", subdir_path) + iterate_folders(subdir_path) + + +def clone_folder_structure(root_src, root_dst, verbose=False): + src_path_list = iterate_folders(root_src) + src_relpath_list = [os.path.relpath(x, root_src) for x in src_path_list] + + os.makedirs(root_dst, exist_ok=True) + dst_path_list = [os.path.join(root_dst, x) for x in src_relpath_list] + for folder_path in dst_path_list: + os.makedirs(folder_path, exist_ok=True) + if verbose: + print(f"Create folder: '{folder_path}'") + + +def count_files(root, suffix=".mp4"): + files_list = iterate_files(root) + cnt = len([x for x in files_list if x.endswith(suffix)]) + return cnt + + +def check_mp4_integrity(file_path, verbose=True, logger=None): + try: + VideoFileClip(file_path) + if verbose: + print_log(f"The MP4 file '{file_path}' is intact.", logger=logger) + return True + except Exception as e: + if verbose: + print_log(f"Error: {e}", logger=logger) + print_log(f"The MP4 file '{file_path}' is not intact.", logger=logger) + return False + + +def count_frames(video_path): + cap = cv2.VideoCapture(video_path) + + if not cap.isOpened(): + print(f"Error: Could not open video file '{video_path}'") + return + + total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT)) + print(f"Total frames in the video '{video_path}': {total_frames}") + + cap.release() + + +def split_video( + sample_path, + scene_list, + save_dir, + target_fps=30, + min_seconds=1, + max_seconds=10, + shorter_size=512, + verbose=False, + logger=None, +): + FFMPEG_PATH = get_ffmpeg_exe() + + save_path_list = [] + for idx, scene in enumerate(scene_list): + s, t = scene # FrameTimecode + fps = s.framerate + max_duration = FrameTimecode(timecode="00:00:00", fps=fps) + max_duration.frame_num = round(fps * max_seconds) + duration = min(max_duration, t - s) + if duration.get_frames() < round(min_seconds * fps): + continue + + # save path + fname = os.path.basename(sample_path) + fname_wo_ext = os.path.splitext(fname)[0] + # TODO: fname pattern + save_path = os.path.join(save_dir, f"{fname_wo_ext}_scene-{idx}.mp4") + + # ffmpeg cmd + cmd = [FFMPEG_PATH] + + # Only show ffmpeg output for the first call, which will display any + # errors if it fails, and then break the loop. We only show error messages + # for the remaining calls. + # cmd += ['-v', 'error'] + + # input path + cmd += ["-i", sample_path] + + # clip to cut + cmd += ["-nostdin", "-y", "-ss", str(s.get_seconds()), "-t", str(duration.get_seconds())] + + # target fps + # cmd += ['-vf', 'select=mod(n\,2)'] + cmd += ["-r", f"{target_fps}"] + + # aspect ratio + cmd += ["-vf", f"scale='if(gt(iw,ih),-2,{shorter_size})':'if(gt(iw,ih),{shorter_size},-2)'"] + # cmd += ['-vf', f"scale='if(gt(iw,ih),{shorter_size},trunc(ow/a/2)*2)':-2"] + + cmd += ["-map", "0", save_path] + + proc = subprocess.Popen(cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT) + stdout, stderr = proc.communicate() + if verbose: + stdout = stdout.decode("utf-8") + print_log(stdout, logger=logger) + + save_path_list.append(sample_path) + print_log(f"Video clip saved to '{save_path}'", logger=logger) + + return save_path_list diff --git a/train.py b/train.py deleted file mode 100644 index 2d48f636..00000000 --- a/train.py +++ /dev/null @@ -1,278 +0,0 @@ -# Copyright (c) Meta Platforms, Inc. and affiliates. -# All rights reserved. - -# This source code is licensed under the license found in the -# LICENSE file in the root directory of this source tree. - -""" -A minimal training script for DiT using PyTorch DDP. -""" -import argparse -import os -from functools import partial - -import torch -import torch.distributed as dist -from colossalai import launch_from_torch -from colossalai.accelerator import get_accelerator -from colossalai.booster import Booster -from colossalai.cluster import DistCoordinator -from colossalai.logging import get_dist_logger -from colossalai.nn.lr_scheduler import CosineAnnealingLR -from colossalai.nn.optimizer import HybridAdam -from colossalai.utils import get_current_device -from torch.utils.tensorboard import SummaryWriter -from tqdm import tqdm - -from open_sora.diffusion import create_diffusion -from open_sora.modeling import DiT_models -from open_sora.modeling.dit import SUPPORTED_MODEL_ARCH, SUPPORTED_SEQ_PARALLEL_MODES -from open_sora.utils.data import ( - create_video_compressor, - load_datasets, - make_batch, - preprocess_batch, -) -from open_sora.utils.plugin import ZeroSeqParallelPlugin - -################################################################################# -# Training Helper Functions # -################################################################################# - - -def configure_backends(): - # the first flag below was False when we tested this script but True makes A100 training a lot faster: - torch.backends.cuda.matmul.allow_tf32 = True - torch.backends.cudnn.allow_tf32 = True - - -@torch.no_grad() -def update_ema(ema_model, model, decay=0.9999): - """ - Step the EMA model towards the current model. - """ - - for ema_p, p in zip(ema_model.parameters(), model.parameters()): - ema_p.mul_(decay).add_(p.data, alpha=1 - decay) - - -def requires_grad(model, flag=True): - """ - Set requires_grad flag for all parameters in a model. - """ - for p in model.parameters(): - p.requires_grad = flag - - -def all_reduce_mean(tensor: torch.Tensor) -> torch.Tensor: - dist.all_reduce(tensor=tensor, op=dist.ReduceOp.SUM) - tensor.div_(dist.get_world_size()) - return tensor - - -def save_checkpoints(booster, model, optimizer, ema, save_path, coordinator): - os.makedirs(save_path, exist_ok=True) - booster.save_model(model, os.path.join(save_path, "model"), shard=False) - booster.save_optimizer(optimizer, os.path.join(save_path, "optimizer"), shard=True) - if coordinator.is_master(): - ema_state_dict = ema.state_dict() - for k, v in ema_state_dict.items(): - ema_state_dict[k] = v.cpu() - torch.save(ema_state_dict, os.path.join(save_path, "ema.pt")) - dist.barrier() - - -################################################################################# -# Training Loop # -################################################################################# - - -def main(args): - """ - Trains a new DiT model. - """ - # Step 1: init distributed environment - launch_from_torch({}) - coordinator = DistCoordinator() - logger = get_dist_logger() - configure_backends() - - # Step 2: set up acceleration plugins - plugin = ZeroSeqParallelPlugin(sp_size=args.sp_size, stage=2, precision="fp16") - booster = Booster(plugin=plugin) - - if coordinator.is_master(): - os.makedirs(args.checkpoint_dir, exist_ok=True) - os.makedirs(args.tensorboard_dir, exist_ok=True) - writer = SummaryWriter(args.tensorboard_dir) - - # Step 3: Create video compressor - video_compressor = create_video_compressor(args.compressor) - model_kwargs = { - "in_channels": video_compressor.out_channels, - "seq_parallel_group": plugin.sp_group, - "seq_parallel_mode": args.sp_mode, - "seq_parallel_overlap": args.sp_overlap, - "model_arch": args.model_arch, - } - - # Step 4: Create DiT and EMA - model = DiT_models[args.model](**model_kwargs).to(get_current_device()) - patch_size = model.patch_size - ema = DiT_models[args.model](**model_kwargs).to(get_current_device()) - update_ema(ema, model, decay=0) - requires_grad(ema, False) - - model.train() # important! This enables embedding dropout for classifier-free guidance - ema.eval() # EMA model should always be in eval mode - - # configure gradient checkpointing - if args.grad_checkpoint: - model.enable_gradient_checkpointing() - - # Step 5: create diffusion pipeline - diffusion = create_diffusion( - timestep_respacing="" - ) # default: 1000 steps, linear noise schedule - - # Step 6: setup optimizer (we used default Adam betas=(0.9, 0.999) and a constant learning rate of 1e-4 in our paper): - opt = HybridAdam(model.parameters(), lr=args.lr, weight_decay=0) - - # Step 7: Setup dataloader - dataset = load_datasets(args.dataset) - dataloader = plugin.prepare_dataloader( - dataset, - batch_size=args.batch_size, - collate_fn=partial( - make_batch, - video_dir=args.video_dir, - pad_to_multiple=args.sp_size, - use_pooled_text=args.model_arch == "adaln", - ), - shuffle=True, - drop_last=True, - ) - lr_scheduler = CosineAnnealingLR( - opt, args.epochs * len(dataloader) // args.accumulation_steps - ) - logger.info(f"Dataset contains {len(dataset)} samples", ranks=[0]) - - # Step 8: setup booster - model, opt, _, dataloader, lr_scheduler = booster.boost( - model, opt, dataloader=dataloader, lr_scheduler=lr_scheduler - ) - if args.load_model is not None: - booster.load_model(model, args.load_model) - if args.load_optimizer is not None: - booster.load_optimizer(opt, args.load_optimizer) - logger.info( - f"Booster init max device memory: {get_accelerator().max_memory_allocated() / 1024 ** 2:.2f} MB", - ranks=[0], - ) - - # Step 9: Train - num_steps_per_epoch = len(dataloader) // args.accumulation_steps - - for epoch in range(args.epochs): - dataloader.sampler.set_epoch(epoch) - with tqdm( - desc=f"Epoch {epoch}", - disable=not coordinator.is_master(), - total=num_steps_per_epoch, - ) as pbar: - total_loss = torch.tensor(0.0, device=get_current_device()) - for step, batch in enumerate(dataloader): - batch = preprocess_batch( - batch, - patch_size, - video_compressor, - pad_to_multiple=args.sp_size, - model_arch=args.model_arch, - ) - video_inputs = batch.pop("video_latent_states") - mask = batch.pop("video_padding_mask") - t = torch.randint( - 0, - diffusion.num_timesteps, - (video_inputs.shape[0],), - device=video_inputs.device, - ) - loss_dict = diffusion.training_losses( - model, video_inputs, t, batch, mask=mask - ) - loss = loss_dict["loss"].mean() / args.accumulation_steps - total_loss.add_(loss.data) - booster.backward(loss, opt) - - if (step + 1) % args.accumulation_steps == 0: - opt.step() - opt.zero_grad() - lr_scheduler.step() - update_ema(ema, model) - - all_reduce_mean(total_loss) - pbar.set_postfix({"Loss": f"{total_loss.item():.4f}"}) - if coordinator.is_master(): - global_step = (epoch * num_steps_per_epoch) + ( - step + 1 - ) // args.accumulation_steps - writer.add_scalar( - tag="Loss", - scalar_value=total_loss.item(), - global_step=global_step, - ) - pbar.update() - total_loss.zero_() - - # Save DiT checkpoint: - if args.save_interval > 0 and ( - (step + 1) % (args.save_interval * args.accumulation_steps) == 0 - or (step + 1) == len(dataloader) - ): - save_path = os.path.join( - args.checkpoint_dir, f"epoch-{epoch}-step-{step}" - ) - save_checkpoints(booster, model, opt, ema, save_path, coordinator) - logger.info(f"Saved checkpoint to {save_path}", ranks=[0]) - - get_accelerator().empty_cache() - final_save_path = os.path.join(args.checkpoint_dir, "final") - save_checkpoints(booster, model, opt, ema, final_save_path, coordinator) - logger.info(f"Saved checkpoint to {final_save_path}", ranks=[0]) - logger.info( - f"Training complete, max device memory: {get_accelerator().max_memory_allocated() / 1024 ** 2:.2f} MB", - ranks=[0], - ) - - -if __name__ == "__main__": - # Default args here will train DiT-XL/2 with the hyperparameters we used in our paper (except training iters). - parser = argparse.ArgumentParser() - parser.add_argument( - "-m", "--model", type=str, choices=list(DiT_models.keys()), default="DiT-S/8" - ) - parser.add_argument( - "-x", "--model_arch", choices=SUPPORTED_MODEL_ARCH, default="cross-attn" - ) - parser.add_argument("-d", "--dataset", nargs="+", default=[]) - parser.add_argument("-v", "--video_dir", type=str, required=True) - parser.add_argument("-e", "--epochs", type=int, default=10) - parser.add_argument("-b", "--batch_size", type=int, default=4) - parser.add_argument("-g", "--grad_checkpoint", action="store_true", default=False) - parser.add_argument("-a", "--accumulation_steps", default=1, type=int) - parser.add_argument("--sp_size", type=int, default=1) - parser.add_argument( - "--sp_mode", type=str, default="ulysses", choices=SUPPORTED_SEQ_PARALLEL_MODES - ) - parser.add_argument("--sp_overlap", action="store_true", default=False) - parser.add_argument("--lr", type=float, default=1e-4) - parser.add_argument("--save_interval", type=int, default=20) - parser.add_argument("--checkpoint_dir", type=str, default="checkpoints") - parser.add_argument("--tensorboard_dir", type=str, default="runs") - parser.add_argument( - "-c", "--compressor", choices=["raw", "vqvae", "vae"], default="raw" - ) - parser.add_argument("--load_model", default=None) - parser.add_argument("--load_optimizer", default=None) - args = parser.parse_args() - main(args) diff --git a/version.txt b/version.txt deleted file mode 100644 index 8acdd82b..00000000 --- a/version.txt +++ /dev/null @@ -1 +0,0 @@ -0.0.1