Skip to content

Releases: moabitcoin/ig65m-pytorch

v1.1.1

16 Oct 08:04
@daniel-j-h daniel-j-h
v1.1.1
686f105
Compare
Choose a tag to compare
v1.1.1 Latest
Latest

Bugfix release; changes

  • Fixed memory usage issue in extract tool
Assets 2
Loading

v1.1.0

14 Oct 20:41
@daniel-j-h daniel-j-h
v1.1.0
14e0a01
Compare
Choose a tag to compare
v1.1.0

This release bring convenience integrations to make
simple tasks such as extracting video features easy.

Changes

  • nvidia-docker based GPU Docker integration; we can
    run self-contained reproducible images on GPUs now

  • Provides 8 and 32 frame models and weights trained
    on IG65M and optionally fine-tuned on Kinetics-400:
    r2plus1d_34_32_ig65m, r2plus1d_34_32_kinetics,
    r2plus1d_34_8_ig65m, r2plus1d_34_8_kinetics

  • Adds Travis CI/CD integration for integration tests

  • Adds Docker registry integration allowing Travis to
    build and push pre-built Docker images on success

  • Upgrades to latest PyTorch 1.3 and torchvision 0.4.1

  • Provides easy to use video feature extraction tool

  • Provides semcode tool visualizing features over time

PyTorch Hub integration can be used from code as follows:

>>> import torch
>>>
>>> torch.hub.list("moabitcoin/ig65m-pytorch")
['r2plus1d_34_32_ig65m', 'r2plus1d_34_32_kinetics', 'r2plus1d_34_8_ig65m', 'r2plus1d_34_8_kinetics']
>>>
>>> model = torch.hub.load("moabitcoin/ig65m-pytorch", "r2plus1d_34_32_ig65m", num_classes=359, pretrained=True)

Pre-build Docker images for CPUs and GPUs (see all tags) can be used as follows:

docker run moabitcoin/ig65m-pytorch:latest-cpu --help

docker run --ipc=host -v $PWD:/data moabitcoin/ig65m-pytorch:latest-gpu \
    extract /data/myvideo.mp4 /data/myfeatures.npy

docker run --runtime=nvidia --ipc=host -v $PWD:/data moabitcoin/ig65m-pytorch:latest-gpu \
    extract /data/myvideo.mp4 /data/myfeatures.npy
🐌 Running on CPU(s)
100%|███████████████████████████████████████████████████████████| 2/2 [00:16<00:00,  8.27s/it]
🍪 Done

extract: video feature extraction running in 🐌-mode on my year 2015 laptop

Semcode for Primer trailer
semcode: visualizing clip features over time for the Primer movie trailer; time goes top to bottom

Please provide feedback (good or bad 🍪) by means of opening issues or pull requests.

Done by the lovely folks from Berlin

Thanks and happy hacking!

Assets 2
Loading

v1.0.0

27 Sep 17:28
@daniel-j-h daniel-j-h
v1.0.0
a783234
Compare
Choose a tag to compare
v1.0.0

This is the initial release providing

  • A conversion tool for R(2+1)D 34-layer model weights
  • The PyTorch model architectures in sync with official Caffe2 models
  • Ported pre-trained IG-65M & fine-tuned on Kinetics model weights in PyTorch
  • A feature extraction/prediction tool providing an usage example on videos

Please provide feedback (good or bad 🍪) by means of opening issues or pull requests.

Done by the lovely folks from Berlin

Thanks and happy hacking!

Assets 10
Loading