small question about lr_scheduler

[huixiancheng]

Thanks for the opensource code！！
Could you tell me the meaning of metric in lr step?

T2T-ViT/main.py

Lines 577 to 579 in f436fe4

	if lr_scheduler is not None:
	# step LR for next epoch
	lr_scheduler.step(epoch + 1, eval_metrics[eval_metric])

T2T-ViT/main.py

Lines 688 to 689 in f436fe4

	if lr_scheduler is not None:
	lr_scheduler.step_update(num_updates=num_updates, metric=losses_m.avg)

In my understanding.Look like in timm it's don't have special meaning.

[yuanli2333]

Hi, lr_scheduler is the method to set up lr decay in training, such as multi-step decay and cosine decay. We use cosine decay in this work.

[huixiancheng]

I know it's cosine_schedule with warm up.
But the ques is about update the ls after a epoch or batch(iter), there are two paras: epoch/num_updates and metric.
Since I don't use timm before,I cann't understand the meaning of Para:metric
In timm it don't have special meaning too. As the code show.

[yuanli2333]

You need update ls in each epoch to tell it current epoch, which is the ls usage in torch.