2021-12-04 15:18:50 +08:00
|
|
|
|
mindspore.nn.DynamicLossScaleUpdateCell
|
|
|
|
|
|
=======================================
|
|
|
|
|
|
|
|
|
|
|
|
.. py:class:: mindspore.nn.DynamicLossScaleUpdateCell(loss_scale_value, scale_factor, scale_window)
|
|
|
|
|
|
|
|
|
|
|
|
用于动态地更新梯度放大系数(loss scale)的神经元。
|
|
|
|
|
|
|
|
|
|
|
|
使用梯度放大功能进行训练时,初始梯度放大系数值为 `loss_scale_value`。在每个训练步骤中,当出现溢出时,通过计算公式 `loss_scale`/`scale_factor` 减小梯度放大系数。如果连续 `scale_window` 步(step)未溢出,则将通过 `loss_scale` * `scale_factor` 增大梯度放大系数。
|
|
|
|
|
|
|
|
|
|
|
|
该类是 :class:`mindspore.nn.DynamicLossScaleManager` 的 `get_update_cell` 方法的返回值。训练过程中,类 :class:`mindspore.TrainOneStepWithLossScaleCell` 会调用该Cell来更新梯度放大系数。
|
|
|
|
|
|
|
2021-12-05 16:06:57 +08:00
|
|
|
|
**参数:**
|
2021-12-04 15:18:50 +08:00
|
|
|
|
|
2021-12-04 20:36:47 +08:00
|
|
|
|
- **loss_scale_value** (float) - 初始的梯度放大系数。
|
|
|
|
|
|
- **scale_factor** (int) - 增减系数。
|
|
|
|
|
|
- **scale_window** (int) - 未溢出时,增大梯度放大系数的最大连续训练步数。
|
2021-12-04 15:18:50 +08:00
|
|
|
|
|
|
|
|
|
|
**输入:**
|
|
|
|
|
|
|
2021-12-04 20:36:47 +08:00
|
|
|
|
- **loss_scale** (Tensor) - 训练期间的梯度放大系数,shape为 :math:`()`。
|
|
|
|
|
|
- **overflow** (bool) - 是否发生溢出。
|
2021-12-04 15:18:50 +08:00
|
|
|
|
|
|
|
|
|
|
**输出:**
|
|
|
|
|
|
|
|
|
|
|
|
Bool,即输入 `overflow` 。
|
|
|
|
|
|
|
|
|
|
|
|
**支持平台:**
|
|
|
|
|
|
|
|
|
|
|
|
``Ascend`` ``GPU``
|
|
|
|
|
|
|
2021-12-05 16:06:57 +08:00
|
|
|
|
**样例:**
|
2021-12-04 15:18:50 +08:00
|
|
|
|
|
|
|
|
|
|
>>> import numpy as np
|
|
|
|
|
|
>>> from mindspore import Tensor, Parameter, nn
|
|
|
|
|
|
>>> import mindspore.ops as ops
|
|
|
|
|
|
>>>
|
|
|
|
|
|
>>> class Net(nn.Cell):
|
|
|
|
|
|
... def __init__(self, in_features, out_features):
|
|
|
|
|
|
... super(Net, self).__init__()
|
|
|
|
|
|
... self.weight = Parameter(Tensor(np.ones([in_features, out_features]).astype(np.float32)),
|
|
|
|
|
|
... name='weight')
|
|
|
|
|
|
... self.matmul = ops.MatMul()
|
|
|
|
|
|
...
|
|
|
|
|
|
... def construct(self, x):
|
|
|
|
|
|
... output = self.matmul(x, self.weight)
|
|
|
|
|
|
... return output
|
|
|
|
|
|
...
|
|
|
|
|
|
>>> in_features, out_features = 16, 10
|
|
|
|
|
|
>>> net = Net(in_features, out_features)
|
|
|
|
|
|
>>> loss = nn.MSELoss()
|
|
|
|
|
|
>>> optimizer = nn.Momentum(net.trainable_params(), learning_rate=0.1, momentum=0.9)
|
|
|
|
|
|
>>> net_with_loss = nn.WithLossCell(net, loss)
|
|
|
|
|
|
>>> manager = nn.DynamicLossScaleUpdateCell(loss_scale_value=2**12, scale_factor=2, scale_window=1000)
|
|
|
|
|
|
>>> train_network = nn.TrainOneStepWithLossScaleCell(net_with_loss, optimizer, scale_sense=manager)
|
|
|
|
|
|
>>> input = Tensor(np.ones([out_features, in_features]), mindspore.float32)
|
|
|
|
|
|
>>> labels = Tensor(np.ones([out_features,]), mindspore.float32)
|
|
|
|
|
|
>>> output = train_network(input, labels)
|
2021-12-04 20:36:47 +08:00
|
|
|
|
|
2021-12-04 15:18:50 +08:00
|
|
|
|
|
|
|
|
|
|
.. py:method:: get_loss_scale()
|
|
|
|
|
|
|
|
|
|
|
|
获取当前梯度放大系数。
|