Python tf.grad_pass_through用法及代碼示例

使用 f 中提供的前向行為創建 grad-pass-through 操作。

用法

tf.grad_pass_through(
    f
)

使用此函數包裝任何操作，保持其在前向傳遞中的行為，但將反向圖中的原始操作替換為一個標識。例如：

x = tf.Variable(1.0, name="x")
z = tf.Variable(3.0, name="z")

with tf.GradientTape() as tape:
  # y will evaluate to 9.0
  y = tf.grad_pass_through(x.assign)(z**2)
# grads will evaluate to 6.0
grads = tape.gradient(y, z)

另一個例子是'differentiable' 移動平均近似，其中允許梯度流入最後一個輸入移動平均的值，但移動平均仍用於前向傳播：

x = ... # Some scalar value
# A moving average object, we don't need to know how this is implemented
moving_average = MovingAverage()
with backprop.GradientTape() as tape:
  # mavg_x will evaluate to the current running average value
  mavg_x = tf.grad_pass_through(moving_average)(x)
grads = tape.gradient(mavg_x, x) # grads will evaluate to 1.0