Official code for the ECAI 2024 paper, "Reset It and Forget It: Relearning Last-Layer Weights Improves Continual and Transfer Learning", by Lapo Frati, Neil Traft, Jeff Clune, and Nick Cheney. (ECAI paper, arXiv preprint)
This reimplements the OML algorithm (paper) and the ANML algorithm (paper, code) in PyTorch using the higher framework.
The training of a few-shot meta-learned model like ANML consists of:
- Pre-train: the model is pre-trained. This can consist of:
- (Meta-)ASB: repeated episodes of sequential learning, followed by batch learning.
- Standard Mini-Batch i.i.d. Training
- Evaluation (Deployment/Transfer): The model is transferred to a new set of classes which do not exist in the training set.
At this stage, we also have multiple options:
- Continual Transfer: The model is trained on the new classes one-image-at-a-time. Classes are ordered sequentially and no example is seen twice.
- Standard Transfer: The model is transferred to the new classes using standard fine-tuning on batches of data.
The code for these phases can be found inside anml.py or iid.py. Refer to the sections below for how to run.
- Python >=3.5
- PyTorch >=1.6.0
- higher
- numpy
- scikit-learn
- tqdm
To get you set up faster, we support creating a Python environment directly from dependency files.
If you have Anaconda or Miniconda, navigate to the higherANML repo and run:
$ conda env create -n higherANML -f environment.yml
$ conda activate higherANMLIf you want to name the environment differently, replace higherANML in the above lines with your custom name.
This file can be re-generated by: (1) running conda env export --from-history > environment.yml, (2) manually deleting
the prefix: field, (3) adding the necessary pip packages.
To use classic Python tooling, navigate to the higherANML repo and run:
$ python -m venv .venv
$ source .venv/bin/activate
$ pip install -r requirements.txtThis assumes you want to store the virtual environment locally within the repo, as is common practice with venv. If
you have errors and do not have the latest pip, try upgrading to see if that fixes the issue: python -m pip install --upgrade pip.
This file should be manually kept in sync when new dependencies are added.
We do not currently support pipenv because this author could not get it to install PyTorch correctly.
But if you get a working Pipfile please do contribute!
Pre-training can be performed using train_anml.py or train_iid.py. The setup can be specified via config file, and
some variables can also be overridden on the command line. Consider the differences between the following configs to
gain an overview of how to edit the config file.
NOTE: These commands will automatically log the training via Weights & Biases unless explicitly disabled.
- To train a setup equivalent to OML:
python train_anml.py -c configs/train-omni-oml.yml. - To train a setup equivalent to ANML:
python train_anml.py -c configs/train-omni-anml.yml. - To train our Meta-ASB on Omniglot (ANML minus the neuromodulation):
python train_anml.py -c configs/train-omni-sanml.yml. - To train our ASB (here called
seqepfor "Sequential Episodic") on Omniglot:python train_anml.py -c configs/train-omni-seqep-sanml.yml. - To train using standard i.i.d. (using Adam) on Omniglot:
python train_iid.py -c configs/train-omni-iid-sanml.yml.- To add zapping once every epoch during i.i.d. training, change
lobo_rateto1. - You can specify any other optimizer available in PyTorch. See optimization.py.
- To add zapping once every epoch during i.i.d. training, change
| Number of tasks | Accuracy | Eval LR |
|---|---|---|
| 10 | 0.90 | 0.0010 |
| 50 | 0.89 | 0.0010 |
| 75 | 0.86 | 0.0010 |
| 100 | 0.85 | 0.0010 |
| 200 | 0.80 | 0.0008 |
| 300 | 0.76 | 0.0008 |
| 400 | 0.72 | 0.0008 |
| 600 | 0.65 | 0.0008 |
Evaluation can be performed using
python eval_anml.py -c configs/eval-omni-anml.yml --model trained_anmls/256_112_2304_ANML-29999.pthDuring evaluation the model is quite sensitive to learning rates used, the data shown in the above table has been gathered sweeping several learning rates over multiple repeats.
If you have access to a Slurm cluster, and you are using Slacc, then everything from train to test can be automated in a single pipeline.
Here is an example command to train and test a model:
python launch_train.py -c configs/train-omni-sanml.yml --project my-wandb-project
Everything about the training and evaluation is defined by this single config file, configs/train-omni-sanml.yml
(which also references other config files). You should see output like this when launching:
wandb: Currently logged in as: ntraft (use `wandb login --relogin` to force relogin)
wandb: wandb version 0.12.17 is available! To upgrade, please run:
wandb: $ pip install wandb --upgrade
wandb: Tracking run with wandb version 0.12.10
wandb: Syncing run whole-shadow-15297
wandb: ⭐️ View project at https://wandb.ai/ntraft/anml-sanml-oml
wandb: 🚀 View run at https://wandb.ai/ntraft/anml-sanml-oml/runs/29ox9s3q
wandb: Run data is saved locally in /gpfs2/scratch/ntraft/Development/higherANML/wandb/run-20220527_142930-29ox9s3q
wandb: Run `wandb offline` to turn off syncing.
Running command: launcher dggpu /gpfs2/scratch/ntraft/Development/higherANML/train_anml.py --config /gpfs2/scratch/ntraft/Development/higherANML/experiments/anml-sanml-oml/whole-shadow-15297/train-config.yml
Submitted batch job 533897
wandb: Waiting for W&B process to finish, PID 71284... (success).
wandb:
wandb: Synced 6 W&B file(s), 0 media file(s), 0 artifact file(s) and 0 other file(s)
wandb: Synced whole-shadow-15297: https://wandb.ai/ntraft/anml-sanml-oml/runs/29ox9s3q
wandb: Find logs at: /gpfs2/scratch/ntraft/Development/higherANML/wandb/run-20220527_142930-29ox9s3q/logs/debug.log
wandb:
Once launched, you can alter any piece of the config (or the code!) and launch another.
Each run will create its own folder in higherANML/experiments/<project-name>/<group-name>, and all configs and results
will be logged under this group name (in the above case, whole-shadow-15297).