Mixup: Beyond Empirical Risk Minimization in PyTorch

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

The results:

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.

Dataset and Model	Acc.
CIFAR 10 no mixup	94.97%
CIFAR 10 mixup	95.53%
CIFAR 100 no mixup	76.53%
CIFAR 100 mixup	77.83%

CIFAR 10 test accuracy evolution

CIFAR 100 test accuracy evolution

Usage

# 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'