Awesome

Online Continual Learning with Maximally Interfered Retrieval (NeurIPS 2019)

Controlled sampling of memories for replay: retrieving the samples which are most interfered, i.e. whose prediction will be most negatively impacted by the foreseen parameters update. </br>

• paper
• summary
• video

(key) Requirements

• Python 3.6
• Pytorch 1.1.0

pip install -r requirements.txt

Structure

├── Scripts 
    ├── ER_experiments.sh                # reproduces Experience Replay (ER) results        
    ├── ER_experiments_miniimagenet.sh   # reproduces ER results on MiniImagenet
    ├── gen_hparam_search.py             # Hyperparameter search for Generative Replay (GEN) 
    ├── gen_reproduce.py                 # reproduces GEN results 
    ├── hybrid_reproduce.sh              # reproduces Hybrid Replay (AE) results
├── VAE           
    ├── ....            # files for the VAE used in GEN
├── buffer.py           # Basic buffer implementation for ER and AE
├── data.py             # DataLoaders
├── er_main.py          # main file for ER
├── gen_main.py         # main file for GEN    
├── hybrid_main.py      # main file for AE
├── mir.py              # retrieval functions for ER, GEN and AE    
├── model.py            # defines the classifiers and the AutoEncoder in AE
├── utils.py

Running Experiments

• ER = Experience Replay baseline
• ER-MIR = Experience Replay + Maximally Interfered Retrieval
• GEN = Generative Replay baseline
• GEN-MIR = Generative Replay + Maximally Interfered Retrieval
• AE = Hybrid Replay baseline
• AE-MIR = Hybrid Replay + Maximally Interfered Retrieval

Experience Replay

ER baseline example: </br>

python er_main.py --method rand_replay --dataset split_cifar10 --mem_size 50

ER-MIR example: </br>

python er_main.py --method mir_replay --dataset split_cifar10 --mem_size 50

Reproduce: </br>

sh Scripts/ER_experiments.sh

Generative Replay

GEN baseline example: </br>

python gen_main.py --method rand_gen --gen_method rand_gen --samples_per_task 1000

GEN-MIR (MIR only on the classifier): </br>

python gen_main.py --method mir_gen --gen_method rand_gen --samples_per_task 1000

GEN-MIR (MIR only on the generator): </br>

python gen_main.py --method rand_gen --gen_method mir_gen --samples_per_task 1000

GEN-MIR: </br>

python gen_main.py --method mir_gen --gen_method mir_gen --samples_per_task 1000

Hyper-parameter search: </br>

python Scripts/gen_hparam_search.py

Reproduce: </br>

python Scripts/gen_reproduce.py

Hybrid Replay

AE baseline example: </br>

python hybrid_main.py --max_loss --mem_size 1000 --buffer_batch_size 100

AE-MIR example: </br>

python hybrid_main.py --mem_size 1000 --buffer_batch_size 100

Reproduce: </br>

sh Scripts/hybrid_reproduce.sh

Logging

Logging is done with Weights & Biases and can be turned on like this: </br> python <method>_main.py --log online

Acknowledgements

We would like to thank authors of the following repositories (from which we borrowed code) for making the code public. </br>

• https://github.com/riannevdberg/sylvester-flows

Cite

@incollection{NIPS2019_9357,
title = {Online Continual Learning with Maximal Interfered Retrieval},
author = {Aljundi, Rahaf and Belilovsky, Eugene and Tuytelaars, Tinne and Charlin, Laurent and Caccia, Massimo and Lin, Min and Page-Caccia, Lucas},
booktitle = {Advances in Neural Information Processing Systems 32},
editor = {H. Wallach and H. Larochelle and A. Beygelzimer and F. d\textquotesingle Alch\'{e}-Buc and E. Fox and R. Garnett},
pages = {11849--11860},
year = {2019},
publisher = {Curran Associates, Inc.},
url = {http://papers.nips.cc/paper/9357-online-continual-learning-with-maximal-interfered-retrieval.pdf}
}

Questions?

For general questions + GEN related questions, contact Massimo </br> For ER related questions, contact Eugene </br> For AE related questions, contact Lucas </br>