Two epochs being trained when one specified

Part 1 (2020)
1

Hi all, I’m working through Fastbook in Colab and when I run fine_tune(1) it seems to run 2 epochs but prints them separately. Below is an image of what I’m talking about, does anyone have any ideas why that’s happening?

Thanks,
Iain

image
image1189×488 45.2 KB

2

fine_tune follows a transfer learning regiment: one epoch is always trained frozen (unless you specify a frozen_epochs=x), with however many epochs you pass in trained unfrozen with an adjustment to your learning rate applied.

3 Likes
3

Ah, so fine_tune handles both the frozen and the unfrozen training! I wrongly assumed that it was simply a replacement for fit_one_cycle and thought there was an issue with the configured environment.

Thanks for clearing that up Zachary!

4

If you check the source code for it, you will find something akin to:

learn.freeze()
learn.fit_one_cycle(frozen_epochs, lr)
learn.unfreeze()
learn.fit_one_cycle(unfrozen_epochs, lr) #this may be modified, can't remember OOTOMH

Jeremy and Sylvain found that usually one epoch is enough to update the unfrozen weights

3 Likes