CODE for paper "LA3: Label-Aware AutoAugment"

code adapted from https://github.com/kakaobrain/fast-autoaugment

#################################################

main code files:

    LA3/search_stage_1.py:
        perform stage 1 (Augmentation Evaluation) search given a pre-trained model, output a file containing the search history to train neural predictors

    LA3/search_stage_2.py:
        train neural predictors with saved search data, construct the final policy

    LA3/train_label_aware.py
        train target networks with searched label-aware policy

    LA3/train_label_aware_batch.py
        train target networks with searched label-aware policy and batch augment

#################################################

searched policy files:

    policies/cifar10_policy.pkl
    policies/cifar100_policy.pkl
    policies/imgnet_policy.pkl

#################################################

Example usage for CIFAR-100 & WRN-40-2:

    search stage 1:

        python LA3/search_stage_1.py -c confs/wresnet40x2_cifar_search.yaml --dataset reduced_cifar100 --dataroot DATAROOT --pretrained_model CIFAR_100_MODEL_PATH

    search stage 2:

        python LA3/search_stage_2.py --data_file reduced_cifar100_wresnet40_2/aug_pg_data.pkl -c confs/wresnet40x2_cifar.yaml --alpha 2.5

    train:

        python LA3/train_class_op.py -c confs/wresnet40x2_cifar_train.yaml --dataset cifar100_class_op --dataroot DATAROOT --cls_policy POLICY_FILE