gradient.py 文件源码

def zero_grad(x):
    """
    Consider an expression constant when computing gradients.

    The expression itself is unaffected, but when its gradient is
    computed, or the gradient of another expression that this
    expression is a subexpression of, it will be backpropagated
    through with a value of zero. In other words, the gradient of
    the expression is truncated to 0.

    :param x: A Theano expression whose gradient should be truncated.

    :return: The expression is returned unmodified, but its gradient
        is now truncated to 0.
    """
    return zero_grad_(x)