When we make prediction with learn.predict('fn.jpg'), does the picture goes through the same transformations as in the Dataloaders, ie Normalization.
It’ll still normalise in the same way but it won’t do any augmentation (ie flip zoom crop etc).
Close! It will still crop. Any cropping becomes center cropping to that size. (If you want to look up exact behavior, check for what happens in specific transforms when split_idx=1)
2 Likes
Thanks for the answer, but where am I supposed to set the split_idx?
You don’t set it. The transforms do for their behaviors. When doing inference you really only want the validation transforms going unless you have a very specific use case