Understanding lr_find() to implement in PyTorch

Hey guys,
I started learning PyTorch recently. I am trying to implement the learning rate finder in PyTorch from scratch. I am not able to understand how it works. I am confused since it’s all interconnected (CLR, SGDR, CosineAnnealing).
To understand it, I tried looking into fastai library but I got lost. In the documentation written for lr_find(), it says that it uses the technique mentioned in CLR paper. As far as I know, CLR uses a triangular learning rate scheduler. Then how do we get this plot of learning rate vs iterations when we use lr_find();

If it uses CLR, then the plot should be similar to that in paper (triangular plot).Also, the lr_find() has 2 parameters - start_lr and end_lr which has default value of 1e-5 and 10 respectively. But in above graph, the highest value of learning rate (on y-axis) is 1.4. Why ?

I think I am confused because all of this is so interconnected. Can anyone clear my confusion on this, regarding where CLR, SGDR, CosineAnnealing are exactly used ? And how does lr_find() use CLR to find optimal learning rate ?

Thanks :blush:

1 Like

Hey @jayeshsaita, I recently implemented the learning rate finder from scratch. Not exactly identical but it works well. I first looked at torch.optim.lr_scheduler.CosineAnnealingLR class and I defined a LRFinder class based on that in such a way that it matched more or less that plot you showed.

CosineAnnealing is used in the train phase. To find the learning rate the idea is to start with a very small learning rate, like 10^{-7} or so, and increase at each iteration as shown in that plot. As the learning rate increase the loss will decrease faster until in flattens and eventually starts increasing again because the learning rate starts being to high.

I hope this helps! :slight_smile:

1 Like

This my be helpful for you

What’s up with Deep Learning optimizers since Adam?