Camvid learning rate

Part 1 (2019)
1

Hi,

Please refer to this

  1. lr_find(learn)

    learn.recorder.plot()
    When we do there is no need to set the LR.

  2. Then we can set at somewhere 10x smaller than the lowest point?
    say lr = 1e-3 OR example is using 3e-3 which is in the middle between top and bottom ? is 1e-3 ok?
    learn.fit_one_cycle(10, slice(lr), pct_start=0.9)
    How do we set 10, slice(lr) and pct_start =0.9? What’s the meaning of these 3 parameters?

  3. learn.unfreeze()
    lrs = slice(lr/400,lr/4)
    learn.fit_one_cycle(12, lrs, pct_start=0.8)
    How do we set the 12, slice(lr/400,lr/4) and pct_start=0.8 here? Thanks.

I read the docs https://docs.fast.ai/callbacks.lr_finder.html#Learning-Rate-Finder but all the explanation is not consistent and I can’t summarize it.

Anyone pls help. Thanks

1 Like