This is the official implementation for Schedule-Robust Continual Learning (Scroll), a continual learning algorithm (for classification) that performs consistently and robustly against any streaming schedules of training data. Our model is based on a meta-learning formulation of continual learning, and is provably schedule-invariant for task-based schedules.
The code provides sample experiment on CIFAR and ImageNet-R datasets. It also includes various existing continual learning algorithms for comparison. The baselines are based on AML repo.
Pre-trained resnet for model initialization can be downloaded from here. Please place the checkpoint in init_model/ folder.
python main.py --mem_size 500 --method scroll --model_type resnet --dataset cifar100
python main.py --mem_size 2000 --method scroll --model_type vit --dataset cifar100
python main.py --mem_size 2000 --method scroll --model_type vit --dataset imagenetr --data_root {dataset_path}