regarding transfer learning in fastai v1 I have a question:
is a model with:
- normalize(imagenet_stats), pretrained=True, and all layers unfrozen (learn.unfreeze())
and a model with:
- normalize(), pretrained=False, and all layers unfrozen (learn.unfreeze())
My results suggest that there is a difference, but I don’t know where there should be a difference since transfer learning is just about using weights from another model…
imagenet_stats will normalize the input pixels using the mean and standard deviation from the imagenet image catalog. If
imagenet_stats is not passed into
normalize() then it will use the images present to normalize the input pixels, so I would not expect two models to be equivalent that use two different sets of normalizing stats.
Thanks for your reply!
The question is more about the pretrained/unfrozen. When I choose to use pretrained=True but re-train ALL layers, in my understanding, it should be the same as when using a not pretrained model where I also train all layers (or in fastAI all layergroups)