Searched refs:scaling_sens (Results 1 – 10 of 10) sorted by relevance
| /third_party/mindspore/mindspore/nn/wrap/ |
| D | loss_scale.py | 318 scaling_sens = self.scale_sense
320 status, scaling_sens = self.start_overflow_check(loss, scaling_sens)
322 scaling_sens_filled = C.ones_like(loss) * F.cast(scaling_sens, F.dtype(loss))
324 grads = self.hyper_map(F.partial(_grad_scale, scaling_sens), grads)
334 return loss, cond, scaling_sens
501 scaling_sens = self.scale_sense
504 scaling_sens_filled = C.ones_like(loss) * F.cast(scaling_sens, F.dtype(loss))
513 …grads = self.hyper_map(F.partial(shard_grad_scale, scaling_sens * self.degree), grads, self.accu_g…
516 … grads = self.hyper_map(F.partial(grad_scale, scaling_sens * self.degree), grads, accu_grads)
525 ret = (loss, overflow, scaling_sens)
|
| /third_party/mindspore/tests/ut/python/parallel/ |
| D | test_loss_scale.py | 88 scaling_sens = self.loss_scale
90 scaling_sens = sens
95 scaling_sens = F.depend(scaling_sens, clear_status)
96 grads = self.grad(self.network, weights)(x, self.cast(scaling_sens, mstype.float32))
110 return (loss, cond, scaling_sens)
|
| D | test_dataset_util.py | 33 full_tensor = _to_full_tensor(elem, device_num, global_rank, scaling_sens=None)
47 full_tensor = _to_full_tensor(elem, device_num, global_rank, scaling_sens=None)
65 full_tensor = _to_full_tensor(elem, device_num, global_rank, scaling_sens=0.1)
|
| /third_party/mindspore/mindspore/train/ |
| D | _utils.py | 147 def _to_tensor(elem, scaling_sens=None): argument
154 if scaling_sens:
155 elem_tuple = tuple(elem) + (Tensor(scaling_sens, mstype.float32),)
160 if scaling_sens:
161 lst.append(Tensor(scaling_sens, mstype.float32))
|
| D | model.py | 327 scaling_sens = 1
329 scaling_sens = self._loss_scale_manager.get_loss_scale()
331 scaling_sens /= self._device_number
332 return scaling_sens
|
| /third_party/mindspore/tests/st/networks/models/bert/src/ |
| D | bert_for_pre_training.py | 400 scaling_sens = self.loss_scale
402 scaling_sens = sens
407 scaling_sens = F.depend(scaling_sens, clear_status)
415 self.cast(scaling_sens,
419 grads = self.hyper_map(F.partial(grad_scale, scaling_sens * self.degree), grads)
436 return (loss, cond, scaling_sens)
|
| /third_party/mindspore/mindspore/boost/ |
| D | boost_cell_wrapper.py | 374 scaling_sens = self.scale_sense
376 status, scaling_sens = self._start_overflow_check(loss, scaling_sens)
378 scaling_sens_filled = C.ones_like(loss) * F.cast(scaling_sens, F.dtype(loss))
380 grads = self.hyper_map(F.partial(_grad_scale, scaling_sens), grads)
394 return loss, cond, scaling_sens
|
| /third_party/mindspore/tests/perf_test/bert/ |
| D | test_bert_train.py | 162 scaling_sens = Tensor(np.ones([1]).astype(np.float32))
163 inputs = load_test_data(batch_size) + (scaling_sens,)
217 scaling_sens = Tensor(np.ones([1]).astype(np.float32))
218 inputs = load_test_data(batch_size) + (scaling_sens,)
|
| /third_party/mindspore/tests/st/networks/models/resnet50/src_thor/ |
| D | model_thor.py | 261 scaling_sens = 1
263 scaling_sens = self._loss_scale_manager.get_loss_scale()
265 scaling_sens /= self._device_number
266 return scaling_sens
514 scaling_sens = self._get_scaling_sens()
515 next_element = tuple(next_element) + (Tensor(scaling_sens, mstype.float32),)
|
| /third_party/mindspore/mindspore/parallel/ |
| D | _utils.py | 114 def _to_full_tensor(elem, global_device_num, global_rank, scaling_sens=None): argument
167 if scaling_sens:
168 lst.append(Tensor(scaling_sens, mstype.float32))
|