Conda is recommended to create an environment:

conda create --name aux-data --file requirements.txt

The default data directory is set to ./data in the main repository directory, this can be changed in the ./datasets/dataset_config,oy file. Same goes for results path, set to ./results as a default argument in the ./main_incremental.py.

CIFAR10 and CIFAR100 will be downloaded when running their corresponding scripts.

The ILSVRC12 dataset is required for ImageNet experiments and the six datasets are required for large domain shift: Oxford Flowers, MIT Indoor Scenes, CUB-200-2011 Birds, Stanford Cars, FGVC Aircraft, Stanford Actions.

For auxiliary dataset for CIFAR and domain shift experiments the tiny-images dataset is required. The ./create_aux.py script needs to be executed to create the auxiliary set for those experiments.

To reproduce all the experiments from our paper simply run all the scripts from the ./scripts directory.