how to use lr warmup in traing stage? about pytorch-widedeep

[jrzaurin]

Hey @seyoulala thanks for your kind words! :)

I see a number of ways of doing it:

1. Using multiple trainers: to each trainer you can pass a lr_scheduler (see here). Therefore you could do something like

model = WideDeep(...)
first_scheduler = torch.optim.lr_scheduler.StepLR(...)
warmup_trainer = Trainer(model, objective, lr_schedulers=first_scheduler)
# At this point the weights of the model are "warm"
trainer = Trainer(model, objective, ...). # with a 2nd scheduler if you wanted

2. Passing a Sequential LR scheduler: in principle, the Trainer accepts any valid pytorch scheduler. Therefore, you could build your own sequential learning rate scheduler and pass it. If this throws an error let me know, because it would be a bug to fix. Personally, I have never tried it (but I will now :) )

3. And the last one, the fit method allows for a series of warm up parameters 😉. At the moment I implemented two routines. Have a look here, to the finetune parameter, aliased as warmup. Also to this example notebook.

Let me know if you need any more info/help

[seyoulala]

thanks ！build my own learning rate scheduler by SequentialLR of pytorch can deal this problem

[jrzaurin]

okay, thanks @seyoulala