flip_gradients.py 文件源码

def __call__(self, x, l=1.0):
        grad_name = "FlipGradient%d" % self.num_calls
        @ops.RegisterGradient(grad_name)
        def _flip_gradients(op, grad):
            return [tf.neg(grad) * l]

        g = tf.get_default_graph()
        with g.gradient_override_map({"Identity": grad_name}):
            y = tf.identity(x)

        self.num_calls += 1
        return y