Online Fast Adaptation and Knowledge Accumulation:

a New Approach to Continual Learning

official code for the paper: https://arxiv.org/abs/2003.05856

NOTE: you can find the version of the repo that can reproduce the paper in the camera_ready branch.
The master branch is lagging behind a bit. I'll fast-forward it + clean it up in a bit.

(key) Requirements

Python 3.6 or 3.7 (past 3.8 you can't get the proper torch version)
Pytorch 1.2 (proper torch version for our torchmeta)

pip install -r requirements.txt

Structure

├── Config
    ├── ...                   # model configuration
├── Data
    ├── omniglot.py           # fetches the dataset      
    ├── tiered_imagenet.py    # fetches the dataset
├── MAML           
    ├── metalearner
        ├── maml.py           # defines the models, in particular their CL strategy
    ├── model.py              # defines the backbone neural networks
    ├── utils.py              # some utils    
├── Utils
    ├── bgd_lib
        ├── ...         # files for BGD      
├── main.py             # main file
├── dataloaders.py      # defines the experiment setting, constructs the dataloaders    
├── args.py             # arguments
├── template.py         # main file template (if you dont want to use pytorch)

Running Experiments

example: run C-MAML (in verbose mode):

python main.py --model_name ours -v

some notable args:

python main.py --prob_statio 0.98 --num_epochs 0 --cl_strategy always_retrain --meta_lr 0.1 --learn_step_size 1 --per_param_step_size 1 -v

to try different baselines in ['online_sgd', 'fine_tuning', 'MetaCOG', 'MetaBGD', 'MAML','ANIL', 'BGD']

python main.py --model_name <baseline_name>

Reproduce the results

See the camera-ready branch

Logging

Logging is done with Weights & Biases and can be turned on like this:
python main.py --wandb <workspace_name>

Reference

@article{caccia2020online,
  title={Online Fast Adaptation and Knowledge Accumulation: a New Approach to Continual Learning},
  author={Caccia, Massimo and Rodriguez, Pau and Ostapenko, Oleksiy and Normandin, Fabrice and Lin, Min and Caccia, Lucas and Laradji, Issam and Rish, Irina and Lacoste, Alexandre and Vazquez, David and Charlin, Laurent},
  journal={NeurIPS},
  year={2020},
  url={https://arxiv.org/abs/2003.05856},
  keywords={Continual-Meta Learning, Setting}
}

TODO

code to reproduce experiments

Acknowledgements

MAML code comes from https://github.com/tristandeleu/pytorch-maml

Contact

massimo.p.caccia at gmail.com

pclucas14/osaka