| 23 | return total_norm |
| 24 | |
| 25 | class NativeScalerWithGradNormCount: |
| 26 | state_dict_key = "amp_scaler" |
| 27 | |
| 28 | def __init__(self): |
| 29 | self._scaler = torch.cuda.amp.GradScaler() |
| 30 | |
| 31 | def __call__(self, loss, optimizer, clip_grad=None, parameters=None, create_graph=False, update_grad=True,retain_graph=False): |
| 32 | self._scaler.scale(loss).backward(create_graph=create_graph, retain_graph=retain_graph) |
| 33 | if update_grad: |
| 34 | if clip_grad is not None: |
| 35 | assert parameters is not None |
| 36 | self._scaler.unscale_(optimizer) # unscale the gradients of optimizer's assigned params in-place |
| 37 | norm = torch.nn.utils.clip_grad_norm_(parameters, clip_grad) |
| 38 | else: |
| 39 | self._scaler.unscale_(optimizer) |
| 40 | norm = ampscaler_get_grad_norm(parameters) |
| 41 | self._scaler.step(optimizer) |
| 42 | self._scaler.update() |
| 43 | else: |
| 44 | norm = None |
| 45 | return norm |
| 46 | |
| 47 | def state_dict(self): |
| 48 | return self._scaler.state_dict() |
| 49 | |
| 50 | def load_state_dict(self, state_dict): |
| 51 | self._scaler.load_state_dict(state_dict) |
| 52 | |
| 53 | |
| 54 | def create_logger(output_dir, dist_rank=0, name=''): |
nothing calls this directly
no outgoing calls
no test coverage detected