Use default mean and std values?

• We should test on a pre-trained model that predicted with default mean and std values does not degrade performances
• We should train a model with default values and check that the performances are still good
• If both tests are ok, we can
  1. remove the compute_mean_std function (https://gitlab.com/teklia/atr/dan/-/blob/3a1fea7376886822f9cdf44d39385ed009d0182c/dan/manager/dataset.py#L276)
  2. load images using torch
  3. use default mean and std parameters with Normalize transformation (from torchvision)