python - 如何在 TensorFlow GradientTape 中使用多个渐变？

Question

按照以下代码连接了 3 个神经网络，我们如何从初始网络中获取两个梯度？第一个梯度有效，但第二个梯度返回 None 张量。似乎它们彼此无关以获得梯度。我该如何解决这个问题？

with tf.GradientTape() as tape1:
    with tf.GradientTape() as tape2:
        output1 = NN_model1(input1, training=True)
        output2 = NN_model2(output1, training=True)
        output3 = NN_model3([input1, output1, output2], training=True)
        loss1 = -tf.math.reduce_mean(output3)
        loss2 = -tf.math.reduce_mean(output2)
    grad1 = tape2.gradient(loss1, NN_model1.trainable_variables)
grad2 = tape1.gradient(loss2, grad1)
optimizer.apply_gradients(zip(grad2, NN_model1.trainable_variables))

Answer 1

我认为正确的做法应该是这样的:

with tf.GradientTape() as tape:
    output1 = NN_model1(input1, training=True)
    output2 = NN_model2(output1, training=True)
    output3 = NN_model3([input1, output1, output2], training=True)
    loss1 = -tf.math.reduce_mean(output3)
    loss2 = -tf.math.reduce_mean(output2)
grad = tape.gradient([loss1, loss2], NN_model1.trainable_variables)
optimizer.apply_gradients(zip(grad, NN_model1.trainable_variables))