Training model on frozen layers and then unfreezing them

In lesson 2 notebook we first learn on the frozen layers learn.fit(lr, 3, cycle_len=1, cycle_mult=2)
But what is the point of that if few lines below we unfreeze all the layers and do the fitting again?
lrs = np.array([lr/9, lr/3, lr]) learn.unfreeze() learn.fit(lrs, 3, cycle_len=1, cycle_mult=2)

Does the model preserve weights from the first fitting (if so is there any way to check where they are?) or it was just for the sake of comparison results for frozen vs unfrozen layers?

Replying to this oldish thread because I had the same question and it came up in search - yes, the model (strictly speaking, the learner) will remember the weights from the first fitting.

1 Like