pytorch-ewc

Unofficial PyTorch implementation of DeepMind's paper Overcoming Catastrophic Forgetting, PNAS 2017.

Results

Continual Learning without EWC (left) and with EWC (right).

Installation

$ git clone https://github.com/kuc2477/pytorch-ewc && cd pytorch-ewc
$ pip install -r requirements.txt

CLI

Implementation CLI is provided by main.py

Usage

$ ./main.py --help
$ usage: EWC PyTorch Implementation [-h] [--hidden-size HIDDEN_SIZE]
                                  [--hidden-layer-num HIDDEN_LAYER_NUM]
                                  [--hidden-dropout-prob HIDDEN_DROPOUT_PROB]
                                  [--input-dropout-prob INPUT_DROPOUT_PROB]
                                  [--task-number TASK_NUMBER]
                                  [--epochs-per-task EPOCHS_PER_TASK]
                                  [--lamda LAMDA] [--lr LR]
                                  [--weight-decay WEIGHT_DECAY]
                                  [--batch-size BATCH_SIZE]
                                  [--test-size TEST_SIZE]
                                  [--fisher-estimation-sample-size FISHER_ESTIMATION_SAMPLE_SIZE]
                                  [--random-seed RANDOM_SEED] [--no-gpus]
                                  [--eval-log-interval EVAL_LOG_INTERVAL]
                                  [--loss-log-interval LOSS_LOG_INTERVAL]
                                  [--consolidate]

optional arguments:
  -h, --help            show this help message and exit
  --hidden-size HIDDEN_SIZE
  --hidden-layer-num HIDDEN_LAYER_NUM
  --hidden-dropout-prob HIDDEN_DROPOUT_PROB
  --input-dropout-prob INPUT_DROPOUT_PROB
  --task-number TASK_NUMBER
  --epochs-per-task EPOCHS_PER_TASK
  --lamda LAMDA
  --lr LR
  --weight-decay WEIGHT_DECAY
  --batch-size BATCH_SIZE
  --test-size TEST_SIZE
  --fisher-estimation-sample-size FISHER_ESTIMATION_SAMPLE_SIZE
  --random-seed RANDOM_SEED
  --no-gpus
  --eval-log-interval EVAL_LOG_INTERVAL
  --loss-log-interval LOSS_LOG_INTERVAL
  --consolidate

Train

$ python -m visdom.server &
$ ./main.py               # Train the network without consolidation.
$ ./main.py --consolidate # Train the network with consolidation.

Update Logs

2019.06.29
- Fixed a critical bug within model.estimate_fisher(): Squared gradients of log-likelihood w.r.t. each layer were mean-reduced over all the dimensions. Now it correctly estimates the Fisher matrix by averaging only over the batch dimension
2019.03.22
- Fixed a critical bug within model.estimate_fisher(): Fisher matrix were being estimated with squared expectation of gradient of log-likelihoods. Now it estimates the Fisher matrix with the expectation of squared gradient of log-likelihood.
- Changed the default optimizer from Adam to SGD
- Migrated the project to PyTorch 1.0.1 and visdom 0.1.8.8

Reference

Overcoming Catastrophic Forgetting, PNAS 2017

Author

Ha Junsoo / @kuc2477 / MIT License

xsx1001 / pytorch-ewc