Fixing Wide ResNet
Correct implementation of Wide ResNets (compared to other found on github).
Top1 Cifar10 for WRN 28x10 dropout=0.3
| Augmentation strategy | Top1 accuracy |
|---|---|
| padding 4 & crop, random horizontal flip | 95.77% |
| + Random Erase | 96.84% |
| + RE+RandAugment | 97.27% |
| + RE+RA+CutMix | 97.28% |