Lower/Upper Bound for LR and Upper Bound decay about adamwr

[mpyrozhok]

Hello, @uyekt
I've refactored scheduler a little bit. max_lr could be adjusted by directly specifying it as a parameter for optimizer, base_lr equivalent is min_lr parameter. Maximum lr scaling and lr policies are implemented using callable objects to allow saving scheduler state between runs.

• parameter eta_on_restart_cb=ReduceMaxLROnRestart(ratio=0.5)) will enable scaling max lr by ratio on each restart. ratio=0.5 will decrease max lr by half.
• parameter eta_on_iteration_cb=ExpReduceMaxLROnIteration(gamma=0.99999)) will enable scaling max lr exponentially on every iteration.
• you can create your own scaling policy by writing a callable object or lambda function and passing it as a parameter. eta_on_restart_cb accepts functions that adjust eta on every restart, eta_on_iteration_cb executes scaler function on every iteration.

The only thing that comes to mind concerning your diverging model without knowing details is that your lr is too high. Does your model converge with constant lr?

[uyekt]

Hi @mpyrozhok,

Thanks a lot! Seems to work like a charm!
It converges with constant LR and with triangular too, but sometimes with triangular it diverges on a restart. (loss suddenly around 12 times highers and it doesn't recover)

I am training a CARN for SISR with L1 Loss and AdamW. I tried using L1-Charbonnier and it doesn't start to diverge anymore after the cosine restart. It seems it has something to do with the loss.
I will have to do more tests this week.