/online_ema

Code for the CoLLAs 2023 paper "Improving Online Continual Learning Performance and Stability With Temporal Ensemble"

Primary LanguagePythonMIT LicenseMIT

Code for "Improving Online Continual Learning Performance and Stability With Temporal Ensemble"

Code for the paper Improving Online Continual Learning Performance and Stability With Temporal Ensemble, Soutif--Cormerais et. al., CoLLAs 2023.

The code is based on Pytorch and Avalanche.

Installation

Install pytorch with conda (Instructions)

install the environment and update it using the environment file

cd avalanchev3
conda env create -f environment.yml
conda env config vars set PYTHONPATH=online_ema.git_path:avalanchev3_path

Change the data directory DATADIR inside toolkit/dataset.py to match the one on your system

Create the results dir

How to use

To run single experiments, find the appropriate config file in config dir an run

python main_noboundaries.py --config config/experiment_config.yml

Add the EMA ensembling and parallel evaluation for more efficient continual evaluation

python main_noboundaries.py --config config/experiment_config.yml --mean_evaluation --parallel_evaluation --eval_every 1

License

Please check the License file listed in this repository.

Cite

@inproceedings{soutifcormerais2023improving,
  title={Improving Online Continual Learning Performance and Stability with Temporal Ensembles},
  author={Soutif--Cormerais, Albin and Carta, Antonio and Van de Weijer, Joost},
  booktitle={Conference on Lifelong Learning Agents},
  pages={828--845},
  year={2023},
  organization={PMLR}
}