Rethinking Experience Replay: a Bag of Tricks for Continual Learning

This code is based on our framework: Mammoth - An Extendible Continual Learning Framework for Pytorch.

To run experiments with the default arguments use python ./utils/main.py --model=<MODEL> --dataset=<DATASET> --buffer_size=<MEM_BUFFER_SIZE> --load_best_args.

Available models:

• sgd: SGD with no countermeasure to catastrophic forgetting (lower bound)
• joint: joint training on the whole dataset (upper bound - not continual)
• agem: A-GEM
• gem: Gradient Episodic Memory for Continual Learning
• hal: Hindisght Anchor Learning
• iCaRL: Incremental Classifier and Representation Learning
• er: naive Experience Replay with no tricks
• er_tricks: Experience Replay equipped with our proposed tricks

Available datasets:

• seq-fmnist: Split Fashion-MNIST (5 tasks, 2 classes per task)
• seq-cifar10: Split CIFAR-10 (5 tasks, 2 classes per task)
• seq-cifar100: Split CIFAR-100 (10 tasks, 10 classes per task)
• seq-core50: CORe50 dataset according to the SIT-NC protocol described here

Best args are provided for the following memory buffer sizes:

• 200 exemplars
• 500 exemplars
• 1000 exemplars