mindspore.ops.stop_gradient

mindspore.ops.stop_gradient(value)[source]

StopGradient is used for eliminating the effect of a value on the gradient, such as truncating the gradient propagation from an output of a function. For more details, please refer to Stop Gradient.

Parameters: value (Any) – The value whose effect on the gradient to be eliminated.
Returns: The same as value.

Supported Platforms:: Ascend GPU CPU

Examples

>>> import mindspore
>>> def f1(x):
...     return x ** 2
>>> x = 3.0
>>> f1(x)
9.0
>>> mindspore.ops.grad(f1)(mindspore.tensor(x))
Tensor(shape=[], dtype=Float32, value= 6)
>>>
>>> # The same function with stop_gradient, return a zero gradient because x is effectively treated as a constant.
>>> def f2(x):
...     return mindspore.ops.stop_gradient(x) ** 2
>>> f2(x)
9.0
>>> mindspore.ops.grad(f2)(mindspore.tensor(x))
Tensor(shape=[], dtype=Float32, value= 0)