ReduceLROnPlateau reduces LR without need

1

Hi,

I have a problem with ReduceLROnPlateau. I initiated it using:
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optim, verbose=True, patience=5)

and then, after each epoch (validation), I perform
scheduler.step(loss_meter_validation.mean)

The problem is, that the scheduler decreases LR even though the loss does not plateau.

train_loss

Can somebody please help me out?

Thank you very much! :blush:

2

Could you print the values of loss_meter_validation.mean for the first steps, which seem to trigger the decrease in the learning rate, please?