Disclaimer: We're currently in the process of refactoring and uploading code to this repository. The current version is not thoroughly tested and may break.
The pipeline for running ScanMix is the following:
For CIFAR:
Step 1:
python simclr.py --config_env configs/env.yml --config_exp configs/pretext/<config_file.yml>
Step 2:
python scan.py --config_env configs/env.yml --config_exp configs/scan/<config_file.yml>
Step 3:
(symmetric) python ScanMix.py --r [0.2/0.5/0.8/0.9] --noise_mode sym --config_env configs/env.yml --config_exp configs/scanmix/<config_file.yml>
(asymmetric) python ScanMix.py --r [0.4/0.49] --noise_mode asym --config_env configs/env.yml --config_exp configs/scanmix/<config_file.yml>
(semantic) python ScanMix.py --r 0.4 --noise_mode semantic_[vgg/resnet/densenet] --config_env configs/env.yml --config_exp configs/scanmix/<config_file.yml>
For Webvision
Step 0:
Use moco-v2 to pre-train an InceptionResNetV2 on Webvision and save the model weights as moco_pretrained/webvision.tar.
Step 1:
python moco.py --dataset webvision --config_env configs/env.yml --config_exp configs/pretext/<config_file.yml>
Step 2:
python scan.py --config_env configs/env.yml --config_exp configs/scan/<config_file.yml>
Step 3:
python ScanMix_parallel.py --config_env configs/env.yml --config_exp configs/scanmix/<config_file.yml> --cudaids 0 1