(self,
loss,
optimizer,
clip_grad=None,
parameters=None,
create_graph=False,
update_grad=True)
| 356 | self._scaler = torch.cuda.amp.GradScaler() |
| 357 | |
| 358 | def __call__(self, |
| 359 | loss, |
| 360 | optimizer, |
| 361 | clip_grad=None, |
| 362 | parameters=None, |
| 363 | create_graph=False, |
| 364 | update_grad=True): |
| 365 | self._scaler.scale(loss).backward(create_graph=create_graph) |
| 366 | if update_grad: |
| 367 | if clip_grad is not None: |
| 368 | assert parameters is not None |
| 369 | self._scaler.unscale_(optimizer) # unscale the gradients of optimizer's assigned params in-place |
| 370 | norm = torch.nn.utils.clip_grad_norm_(parameters, clip_grad) |
| 371 | else: |
| 372 | self._scaler.unscale_(optimizer) |
| 373 | norm = get_grad_norm(parameters) |
| 374 | self._scaler.step(optimizer) |
| 375 | self._scaler.update() |
| 376 | else: |
| 377 | norm = None |
| 378 | return norm |
| 379 | |
| 380 | def state_dict(self): |
| 381 | return self._scaler.state_dict() |
nothing calls this directly
no test coverage detected