PyTorch implementation of AutoAugment
This repository contains code for AutoAugment (only using paper's best policies) based on AutoAugment: Learning Augmentation Policies from Data implemented in PyTorch.
Requirements
- Python 3.6
- PyTorch 1.0
Training
CIFAR-10
WideResNet28-10 baseline on CIFAR-10:
python train.py
WideResNet28-10 +Cutout, AutoAugment on CIFAR-10:
python train.py --cutout True --auto-augment True
Results
CIFAR-10
| Model | Error rate | Loss | Error rate (paper) |
|---|---|---|---|
| WideResNet28-10 baseline | 3.82 | 0.1576 | 3.87 |
| WideResNet28-10 +Cutout | 3.40 | 0.1280 | 3.08 |
| WideResNet28-10 +Cutout, AutoAugment | 2.91 | 0.0994 | 2.68 |
Learning curves of loss and accuracy.
