This is the official reimplementation of FasterAutoAugment (hataya2020a.)
Python>=3.8# Developed on 3.8. It may work with earlier versions.PyTorch==1.5.0# Required bykorniatorchvision==0.6.0kornia==0.3.1homura==2020.07#pip install -U git+https://github.com/moskomule/[email protected]hydra==0.11
Faster AutoAugment is an offline method, so first search then train CNN models.
python search.py [data.name={cifar10,cifar100,svhn}] [...]
This script will save the obtained policy at policy_weight/DATASET_NAME/EPOCH.pt.
python train.py path=PATH_TO_POLICY_WEIGHT [data.name={cifar10,cifar100,svhn}] [model.name={wrn28_2,wrn40_2,wrn28_10}] [...]
When path is not specified, training is executed without policy, which can be used as a baseline.
The codebase here is not exactly the same as the one used in the paper.
For example, this codebase does not include the support for DistributedDataParallel and the custom CutOut kernel.
But, we believe this reimplementation is much simpler, more stable, and easier to extend for your research projects.
@inproceesings{hataya2020a,
title={{Faster AutoAugment: Learning Augmentation Strategies using Backpropagation}},
author={Ryuichiro Hataya and Jan Zdenek and Kazuki Yoshizoe and Hideki Nakayama},
year={2020},
booktitle={ECCV}
}