Repeated Augmented Rehearsal (RAR) for online continual learning

This is the official code repository for Repeated Augmented Rehearsal (NeurIPS 2022).

Requirements

Create a virtual enviroment

virtualenv rar

Activating a virtual environment

source rar/bin/activate

Installing packages

pip install -r requirements.txt

Run commands

bash runs

A test run of Repeated Augmented Rehearsal(RAR) with experience replay can be performed with the following command:

bash run_commands/runs/run_test_rar_er_cifar100.sh

Other experiment commands can be found in the folder of run_commands/runs.

Detailed descriptions of options can be found in general_main.py and utils/argparser

For example:

The number of repeated iteration is set via:

--mem_iters $MEM_ITER

The number of augmentation strength is set via:

--randaug True --randaug_N $RAUG_N  --randaug_M $RAUG_M

Evaluation the results

The results of algorithm outputs will be stored in the folder of results.

The jupyter notebook visualize_results.ipynb is used to visualize and analyze results.

Algorithms

Baselines

• LwF: Learning without forgetting (ECCV, 2016) [Paper]
• AGEM: Averaged Gradient Episodic Memory (ICLR, 2019) [Paper]
• ER: Experience Replay (ICML Workshop, 2019) [Paper]
• ASER: Adversarial Shapley Value Experience Replay(AAAI, 2021) [Paper]
• MIR: Maximally Interfered Retrieval (NeurIPS, 2019) [Paper]
• SCR: Supervised Contrastive Replay (CVPR Workshop, 2021) [Paper]
• DER: Dark Experience Replay (NeurIPS, 2020) [Paper]

Datasets

Online Class Incremental

• Split CIFAR100
• Split Mini-ImageNet
• CORe50-NC
• CLRS-NC (Continual Learning Benchmark for Remote Sensing Image Scene Classification)

Data preparation

• CIFAR100 will be downloaded during the first run
• CORE50 download: source fetch_data_setup.sh
• Mini-ImageNet: Download from https://www.kaggle.com/whitemoon/miniimagenet/download , and place it in datasets/mini_imagenet/
• CLRS: Download from https://github.com/lehaifeng/CLRS

Acknowledgments

Our code is based on :

• SCR/ASER
• Rehearsal Revealed
• DER
• MIR
• AGEM