Hi All,
I am working on a multi-label classification problem which is similar to the lesson 3 planets problem as discussed in the fastai course v3.
I start the training with the initial learning rate value of 1e-1 (0.01). After that, I unfreeze and run learn.lr_find() which gave me the following plot.
From the plot, will slice(1e-6, lr/5) is fine.
Your help will be much appreciated.
Thanks a lot
Bilal
Use max_lr=5e-4 and max_lr=5e-6 and see the validaiton loss curve for some epochs like 5. Whichever gives better results, choose that. I think 8e-4 would give better results.
Try various ranges and see the results for yourself. As Jeremy likes to say, “The answer to the question whether I should do bla, is to do bla and see.”
By various I mean with different start and ends points and size as well.
Happy learning.