Improving Continual Learning by Accurate Gradient Reconstructions of the Past

This repository contains the code for the paper Improving Continual Learning by Accurate Gradient Reconstructions of the Past, TMLR 2023.

Setup

Clone this repo and install the dependencies in requirements.txt in a Python 3.9 environment. For convenience, add the repo's src folder to your Python path.

For running ImageNet experiments, install ffcv following their instructions. Also, run the command below to convert the ImageNet dataset to the FFCV format -- see the ffcv-imagenet repo for details (this requires cloning ffcv-imagenet and setting the environment variables in convert_imagenet_ffcv.sh).

$ python src/data/convert_imagenet_ffcv.sh

Run Experiments

Run the command below to create the experiment config files corresponding to the methods reported in the paper (placed in src/configs).

$ python src/create_exp_config_files.py

Run the command below to launch an experiment with a given config (see src/configs).

$ python src/main.py --config path/to/config.yaml

Plot Results

Run the command below to generate the plots shown in the paper (placed in src/plotting/plots).

$ python src/plotting/plot_results.py

Citation

@article{daxberger2023improving,
    title={Improving Continual Learning by Accurate Gradient Reconstructions of the Past},
    author={Erik Daxberger and Siddharth Swaroop and Kazuki Osawa and Rio Yokota and Richard E Turner and Jos{\'e} Miguel Hern{\'a}ndez-Lobato and Mohammad Emtiyaz Khan},
    journal={Transactions on Machine Learning Research},
    issn={2835-8856},
    year={2023},
    url={https://openreview.net/forum?id=b1fpfCjja1},
}

Acknowledgements