Awesome

pytorch-deep-generative-replay

PyTorch implementation of Continual Learning with Deep Generative Replay, NIPS 2017

Results

Continual Learning on Permutated MNISTs

• Test precisions without replay (left), with exact replay (middle), and with Deep Generative Replay (right).

<img width="250" style="margin: 5px;" src="./arts/permutated-mnist-none.png" /> <img width="250" style="margin: 5px;" src="./arts/permutated-mnist-exact-replay.png" /> <img width="250" style="margin: 5px;" src="./arts/permutated-mnist-generative-replay.png" />

Continual Learning on MNIST-SVHN

• Test precisions without replay (left), with exact replay (middle), and with Deep Generative Replay (right).

<img width="250" style="margin: 5px;" src="./arts/mnist-svhn-none.png" /> <img width="250" style="margin: 5px;" src="./arts/mnist-svhn-exact-replay.png" /> <img width="250" style="margin: 5px;" src="./arts/mnist-svhn-generative-replay.png" />

• Generated samples from the scholar trained without any replay (left) and with Deep Generative Replay (right).

<img width="250" style="margin: 5px;" src="./arts/mnist-svhn-none-sample.jpg" /> <img width="250" style="margin: 5px;" src="./arts/mnist-svhn-generative-replay-sample.jpg" />

• Training scholar's generator without replay (left) and with Deep Generative Replay (right).

<img width="250" style="margin: 5px;" src="./arts/mnist-svhn-none.gif" /> <img width="250" style="margin: 5px;" src="./arts/mnist-svhn-generative-replay.gif" />

Continual Learning on SVHN-MNIST

• Test precisions without replay (left), with exact replay (middle), and with Deep Generative Replay (right).

<img width="250" style="margin: 5px;" src="./arts/svhn-mnist-none.png" /> <img width="250" style="margin: 5px;" src="./arts/svhn-mnist-exact-replay.png" /> <img width="250" style="margin: 5px;" src="./arts/svhn-mnist-generative-replay.png" />

• Generated samples from the scholar trained without replay (left) and with Deep Generative Replay (right).

<img width="250" style="margin: 5px;" src="./arts/svhn-mnist-none-r1-sample.jpg" /> <img width="250" style="margin: 5px;" src="./arts/svhn-mnist-generative-replay-r0.4-sample.jpg" />

• Training scholar's generator without replay (left) and with Deep Generative Replay (right).

<img width="250" style="margin: 5px;" src="./arts/svhn-mnist-none.gif" /> <img width="250" style="margin: 5px;" src="./arts/svhn-mnist-generative-replay.gif" />

Installation

$ git clone https://github.com/kuc2477/pytorch-deep-generative-replay
$ pip install -r pytorch-deep-generative-replay/requirements.txt

Commands

Usage

$ ./main.py --help
$ usage: PyTorch implementation of Deep Generative Replay [-h]
                                                          [--experiment {permutated-mnist,svhn-mnist,mnist-svhn}]
                                                          [--mnist-permutation-number MNIST_PERMUTATION_NUMBER]
                                                          [--mnist-permutation-seed MNIST_PERMUTATION_SEED]
                                                          --replay-mode
                                                          {exact-replay,generative-replay,none}
                                                          [--generator-z-size GENERATOR_Z_SIZE]
                                                          [--generator-c-channel-size GENERATOR_C_CHANNEL_SIZE]
                                                          [--generator-g-channel-size GENERATOR_G_CHANNEL_SIZE]
                                                          [--solver-depth SOLVER_DEPTH]
                                                          [--solver-reducing-layers SOLVER_REDUCING_LAYERS]
                                                          [--solver-channel-size SOLVER_CHANNEL_SIZE]
                                                          [--generator-c-updates-per-g-update GENERATOR_C_UPDATES_PER_G_UPDATE]
                                                          [--generator-iterations GENERATOR_ITERATIONS]
                                                          [--solver-iterations SOLVER_ITERATIONS]
                                                          [--importance-of-new-task IMPORTANCE_OF_NEW_TASK]
                                                          [--lr LR]
                                                          [--weight-decay WEIGHT_DECAY]
                                                          [--batch-size BATCH_SIZE]
                                                          [--test-size TEST_SIZE]
                                                          [--sample-size SAMPLE_SIZE]
                                                          [--image-log-interval IMAGE_LOG_INTERVAL]
                                                          [--eval-log-interval EVAL_LOG_INTERVAL]
                                                          [--loss-log-interval LOSS_LOG_INTERVAL]
                                                          [--checkpoint-dir CHECKPOINT_DIR]
                                                          [--sample-dir SAMPLE_DIR]
                                                          [--no-gpus]
                                                          (--train | --test)

To Run Full Experiments

# Run a visdom server and conduct full experiments
$ python -m visdom.server &
$ ./run_full_experiments

To Run a Single Experiment

# Run a visdom server and conduct a desired experiment
$ python -m visdom.server &
$ ./main.py --train --experiment=[permutated-mnist|svhn-mnist|mnist-svhn] --replay-mode=[exact-replay|generative-replay|none]

To Generate Images from the learned Scholar

$ # Run the command below and visit the "samples" directory
$ ./main.py --test --experiment=[permutated-mnist|svhn-mnist|mnist-svhn] --replay-mode=[exact-replay|generative-replay|none]

Note

• I failed to find the supplementary materials that the authors mentioned in the paper to contain the experimental details. Thus, I arbitrarily chose a 4-convolutional-layer CNN as a solver in this project. If you know where I can find the additional materials, please let me know through the project's Github issue.

Reference

• Continual Learning with Deep Generative Replay, arxiv:1705.08690

Author

Ha Junsoo / @kuc2477 / MIT License