Unofficial implementation of "KeepAugment: A Simple Information-Preserving Data Augmentation Approach". CVPR2021
Model | ResNet-18 | Wide ResNet-28-10 |
---|---|---|
Cutout | 95.6±0.1(paepr) | 96.9±0.1(paepr) |
KeepCutout | 96.2 | 97.2 |
KeepCutout (low resolution) | 96.1 | 97.1 |
KeepCutout (early loss) | 96.2 | TODO |
KeepCutout (low + early) | 96.2 | TODO |
Model | ResNet-18 | Wide ResNet-28-10 |
---|---|---|
AutoAugment | 96.0 | 97.3±0.1(paepr) |
KeepAutoAugment | 96.4 | 97.4 |
KeepAutoAugment (low resolution) | 96.3 | 97.5 |
KeepAutoAugment (early loss) | 96.3 | 97.4 |
KeepAutoAugment (low + early) | 96.5 | TODO |
All results have slight differences(±0.1).
The type of method is
'keep_cutout', 'keep_cutout_low', 'keep_cutout_early', 'keep_cutout_low_early',
'keep_autoaugment', 'keep_autoaugment_low', 'keep_autoaugment_early', 'keep_autoaugment_low_early'.
The type of model is 'resnet', 'wide_resnet'.
python train.py --model resnet --method keep_cutout
baseline & cutout code : https://github.com/uoguelph-mlrg/Cutout
autoaugment code : https://github.com/DeepVoltaire/AutoAugment
saliency map code : https://github.com/sunnynevarekar/pytorch-saliency-maps