search

mims-harvard / UniTS

A unified multi-task time series model.
https://zitniklab.hms.harvard.edu/projects/UniTS/
MIT License
363 stars 45 forks source link

Why there is no loss.backward in training process? #13

Closed freshlu11 closed 2 months ago

freshlu11 commented 2 months ago

I just found one loss.backward() in function: memory_check.

but the aim of this function is to find the max batch size, am I right? So where does the backward process occurs?

freshlu11 commented 2 months ago

Sorry, I have find the update code:

norm_value = scaler(loss*loss_scale, model_optim, clip_grad=max_norm, parameters=self.model.parameters(), create_graph=False, update_grad=((i + 1) % acc_it == 0))