For train, validation and test data I have separate data augmentors. The train data luckily fits in the RAM. So I can compute the mean and std that really helped me to impro