question of scheduler in 2D-Tan

[yxG1005]

In the code, the scheduler is set as

scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, factor=config.TRAIN.FACTOR, patience=config.TRAIN.PATIENCE, verbose=config.VERBOSE)
state['scheduler'].step(-val_state['loss_meter'].avg)

As official document of ReduceLROnPlateau, setting here means learning rate will change (lr = lr * factor) when metric didn't decrease in 20 epochs. (Equals to -val_state['loss_meter'].avg didn't decrease, equals to val loss didn't increase). However, when val loss increase, learning rate should change, this conflicted with the setting in code. I am confused with why here add "-" before val_state['loss_meter'].avg?

[Sy-Zhang]

Hi @yxG1005 , Thank you for pointing this out! You are absolutely right. The use of -val_state['loss_meter'].avg here is indeed a logic error. It conflicts with the default mode='min' of ReduceLROnPlateau.