Closed mark-arm closed 3 years ago
No problem, great work on the repo though, it's nice to see reproduction experiments included (and they showed the normalization didn't matter very much!). If you want others to build on this definitely add a LICENSE file of some kind!
Original code accidentally used the standard deviation as the mean for CIFAR normalization. These values calculated from the torchvision.datasets.CIFAR10 training data.