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

Last update: Dec 14, 2022

Overview

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).

Continual Learning on MNIST-SVHN

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

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

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

Continual Learning on SVHN-MNIST

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

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

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

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

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

Related tags

Overview

pytorch-deep-generative-replay

Results

Continual Learning on Permutated MNISTs

Continual Learning on MNIST-SVHN

Continual Learning on SVHN-MNIST

Installation

Commands

Usage

To Run Full Experiments

To Run a Single Experiment

To Generate Images from the learned Scholar

Note

Reference

Author

Owner

Junsoo Ha

A selection of State Of The Art research papers (and code) on human locomotion (pose + trajectory) prediction (forecasting)

How to Become More Salient? Surfacing Representation Biases of the Saliency Prediction Model

Code to produce syntactic representations that can be used to study syntax processing in the human brain

The Easy-to-use Dialogue Response Selection Toolkit for Researchers

DCT-Mask: Discrete Cosine Transform Mask Representation for Instance Segmentation

Visualizing Yolov5's layers using GradCam

Official implementation for "QS-Attn: Query-Selected Attention for Contrastive Learning in I2I Translation" (CVPR 2022)

The official implementation code of "PlantStereo: A Stereo Matching Benchmark for Plant Surface Dense Reconstruction."

RM Operation can equivalently convert ResNet to VGG, which is better for pruning; and can help RepVGG perform better when the depth is large.

✨风纪委员会自动投票脚本，利用Github Action帮你进行裁决操作（为了让其他风纪委员有案件可判，本程序从中午12点才开始运行，有需要请自己修改运行时间）

HairCLIP: Design Your Hair by Text and Reference Image

Grammar Induction using a Template Tree Approach

Human Action Controller - A human action controller running on different platforms.

Jupyter Dock is a set of Jupyter Notebooks for performing molecular docking protocols interactively, as well as visualizing, converting file formats and analyzing the results.

Official Implementation of "Designing an Encoder for StyleGAN Image Manipulation"

Band-Adaptive Spectral-Spatial Feature Learning Neural Network for Hyperspectral Image Classification

Combining Automatic Labelers and Expert Annotations for Accurate Radiology Report Labeling Using BERT

Create and implement a deep learning library from scratch.

A python library for face detection and features extraction based on mediapipe library

This is an official implementation of the High-Resolution Transformer for Dense Prediction.