Learning rate and learning rate scheduler

[dtdo90]

Dear author,

Thank you for the amazing work!!!

I have a question regarding the choice of learning rate and learing-rate scheduler. I the optimizer.yaml file, it set lr=0.0001
and lr_scheduler as

type: MultiStepLR
milestones: [1000]
gamma: 0.1

Is there a reason for setting such a small learning rate? Furthermore, the scheduler reduces learning rate at steps 1000 (by factor 0.1) which is pretty early in the entire training process?

[lyuwenyu]

Yes, It means we keep fixed lr in training process.