This is an unofficial PyTorch implementation of mixup: Beyond Empirical Risk Minimization. The code is adapted from PyTorch CIFAR.

I only tested using CIFAR 10 and CIFAR 100. The network we used is PreAct ResNet-18. For mixup, we set alpha to be default value 1, meaning we sample the weight uniformly between zero and one. I trained 200 epochs for each setting. The learning rate is 0.1 (iter 1-100), 0.01 (iter 101-150) and 0.001 (iter 151-200). The batch size is 128.

# Train and test CIFAR 10 with mixup.
python main_cifar10.py --mixup --exp='cifar10_nomixup'
# Train and test CIFAR 10 without mixup.
python main_cifar10.py --exp='cifar10_nomixup'
# Train and test CIFAR 100 with mixup.
python main_cifar100.py --mixup --exp='cifar100_mixup'
# Train and test CIFAR 100 without mixup.
python main_cifar100.py --exp='cifar100_nomixup'