Missing input normalization ?

[mazzma12]

Hi,
I might be wrong, but I don't see the place were the input are normalized according to the mean and std of the training set. pred_batch expects tensor out of dataloader, hence already normalized I would say

[drewbo]

@mazzma12 it isn't abstracted correctly right now, but I'm handing normalization in this line, assuming it is the first transformation. Only the first transform is applied right now (which should be updated) and this line could be further improved to operate over a tensor instead of iterating over a list. Let me know if you'd like help with a PR to this effect

[mazzma12]

I understand, but it's not obvious that the first transform is the normalization. For example when you call get_transforms() for generic transformation you end up with a crop pad transform as the only valid transform

RandTransform(tfm=TfmCrop (crop_pad), kwargs={}, p=1.0, resolved={}, do_run=True, is_random=True, use_on_y=True)

[drewbo]

Totally agree, I agree it's a bug just mentioning that it's possible to normalize through the current implementation if it's the only transform (because generally the other transforms aren't applied at inference time)