Does saving a pre-trained frozen model also saves the weights of the frozen layers?

I have a pre-trained resnet34 model. First I train it for a few epochs and save the model. The filesize is 48.2MB.
I then unfreeze the model and continue training for another few epochs. When I save it, the size is 218.3MB.
I then freeze it again and continue training. At the end I save the model, and the filesize is 48.2MB again.
Is it correct that when I save the frozen model, only the weights for the last (unfrozen) layer are saved?
If I load the frozen model in the future, are the weights of all layers except the last ones initiated with the “standard” resnet34 weights?

1 Like

I ran some experiments and came to the conclusion that the weights from both the frozen and unfrozen layers are saved. The difference in file size between a frozen model and unfrozen is because save() by default also saves the optimizer state.


Hey, interesting experiment. Although, I don’t quite follow the explanation. Can you please elaborate about what it means to save the optimizer state? and how that causes the changes in file size?

It’s True. freeze model doesn’t have grad data,but the unfreeze model includes.
you can try this code:


But why do we need to save grad data with models when the model is in unfreezed state? saves the optimizer state by default which contains momentum_buffer for the layers. After unfreezing, this buffer increases so is the .pth file size. I think there is no grad saved. Try this:

learn.fit_one_cycle(1, 1e-2)'stage-1', with_opt=False)
learn.fit_one_cycle(1, slice(1e-5, 1e-3)'stage-2', with_opt=False)

And check the .pth file sizes now.