davda54/sam

Question: Why step_lr?

Closed this issue · 2 comments

Thank you for the great implementation!

Out of curiosity, why have you added your own step_lr scheduler and are not using the cosine annealing scheduler as suggested in the original paper?

It is just a simple example :) I've used my old implementation of WRN, which uses the step_lr scheduler from the original WRN paper. There's certainly a lot of room for improvement if you want to have a better classifier.

Got it. Thank you!