Thanks for sharing! FYI you may find resnet18 or (better still) densenet a good option for that dataset - you shouldn't need so many training images then.