Progressive resizing in fastai v1

Hi guys, just want to confirm here, are the code below still the valid way to do progressive resizing in image classification with fastai v1:

learn.load(‘model-trained-with-size-64’) = get_data(sz=128) # get_data is a little helper function to return an image DataBunch
learn.fit_one_cycle(3, lr, div_factor=40)

lrs = np.array([lr/9,lr/3,lr])
learn.fit_one_cycle(3, lrs, div_factor=40)

I’m asking because my model is currently training on size 128, but I noticed that the losses on the 3 epochs with frozen model (all but last layer) are higher than the losses with unfrozen model with size 64, but better than frozen model with size 64.

This should work yes. A way to be sure would be to load one batch x,y = next(iter( and look at the size of x.

1 Like

Progressive sizing looks OK.
Are you passing in the lrs in the right form? If you pass in a max_lr then fit_one_cycle creates discriminative lrs, see lr_range in docs.
Also beware pct_start for one cycle is not a percentage like 10 in v0, but 0.1 as a fraction for v1. I’d have preferred it stay an integer percent with that name, but not a big deal.
Call learn.recorder.plot_lr(show_moms=True) to check the lr schedule is what you intended.

Size of images in new does get scaled up to the expected size. Thanks

Thanks for the insights. fit_one_cycle does take a list of learning rates as max_lr.
Below line of code comes from right above doc of lr_range: :grinning:

learn.fit_one_cycle(1, max_lr=(1e-4, 1e-3, 1e-2), wd=(1e-4,1e-4,1e-1))

That’s only true if you pass in a slice object, or something listy.

I’m wondering how do I change data for Learner in fp16 mode? If I just set it by setting the .data attribute, it will break saying

Input type (torch.cuda.FloatTensor) and weight type (torch.cuda.HalfTensor) should be the same

If i convert it use .to_fp32(), set .data, then do .to_fp16(), it seems to drop its original weight somehow. The .validate() accuracy changes dramatically before and after .to_fp16.

Just a general question about progressive resizing. Is the main advantage of this technique to speed up training times? Or do we actually get a boost in accuracy directly from this technique?

Also another question. When doing progressive resizing for each size, is it preferable to unfreeze and fine tune for each size? Or should we just unfreeze and fine tune in the training run for the final largest size?

1 Like