/mixup

Implementation of the mixup training method

Primary LanguagePythonBSD 3-Clause "New" or "Revised" LicenseBSD-3-Clause

This repo contains demo reimplementations of the CIFAR-10 training code and the GAN experiment in PyTorch based on the following paper:

Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin and David Lopez-Paz. mixup: Beyond Empirical Risk Minimization. https://arxiv.org/abs/1710.09412

CIFAR-10

The following table shows the median test errors of the last 10 epochs in a 200-epoch training session. (Please refer to Section 3.2 in the paper for details.)

Model weight decay = 1e-4 weight decay = 5e-4
ERM 5.53% 5.18%
mixup 4.24% 4.68%

Generative Adversarial Networks (GAN)

Other implementations

Acknowledgement

The CIFAR-10 reimplementation of mixup is adapted from the pytorch-cifar repository by kuangliu.