Intern project by Geoff French, hosted by Avital Oliver and Tim Salimans
This project explores the use of CowMask for semi-supervised image classification. It achieves competitive results on CIFAR-10 using a 26 layer Wide ResNet with Shake-shake regularization. It also achieves competitive results on ImageNet.
To train a network on CIFAR-10 using 1,000 supervised samples:
python main_semisup.py
The default values for the command line arguments should replicate a single result from our upcoming paper.