$ conda create --name SPR python=3.6

$ pip install -r requirements.txt

$ mkdir data
$ ln -s [MNIST Data Path] data/mnist
$ ln -s [CIFAR10 Data Path] data/cifar10
$ ln -s [CIFAR100 Data Path] data/cifar100
$ ln -s [Webvision Data Path] data/webvision

$ ln -s [log directory path] checkpoints

python main.py --log-dir [log directory path] --c [config file path] --e [episode file path] --override "|" --random_seed [seed]

# e.g. to run mnist symmetric noise 40% experiment,
python main.py --log-dir [log directory path] --c configs/mnist_spr.yaml --e episodes/mnist-split_epc1_a.yaml --override "corruption_percent=0.4";

# e.g. to run cifar10 asymmetric noise 40% experiment,
python main.py --log-dir [log directory path] --c configs/cifar10_spr.yaml --e episodes/cifar10-split_epc1_asym_a.yaml --override "asymmetric_nosie=False|corruption_percent=0.4";

# e.g. to run cifar100 superclass symmetric noise 40% experiment,
python main.py --log-dir [log directory path] --c configs/cifar100_spr.yaml --e episodes/cifar100sup-split_epc1_a.yaml --override "superclass_nosie=True|corruption_percent=0.4";

# e.g. to run mnist symmetric noise 40% experiment,
python meta-main.py --log-dir [log directory path] -c configs/mnist_spr.yaml -e episodes/mnist-split_epc1_a.yaml --random_seed [seed] --override "corruption_percent=0.4" --njobs 10 --jobs_per_gpu 3

# also, you can only train experts for later use by adding an --expert_train_only option.
python meta-main.py --log-dir [log directory path] -c configs/mnist_spr.yaml -e episodes/mnist-split_epc1_a.yaml --random_seed [seed] --override "corruption_percent=0.4" --ngpu 10 --jobs_per_gpu 3 --expert_train_only

## to use the trained experts, set the same [log directory path] and [seed].
python main.py --log-dir [log directory path] --c configs/mnist_spr.yaml --e episodes/mnist-split_epc1_a.yaml --random_seed [seed] --override "corruption_percent=0.4";

@inproceedings{kim-ICCV-2021,
    author    = {Chris Dongjoo Kim and Jinseo Jeong and Sangwoo Moon and Gunhee Kim},
    title     = "{Continual Learning on Noisy Data Streams via Self-Purified Replay}"
    booktitle = {ICCV},
    year      = 2021
}