Deep-Learning-with-TensorFlow-book/ch07-反向传播算法/chain_rule.py

import tensorflow as tf 

# 构建待优化变量
x = tf.constant(1.)
w1 = tf.constant(2.)
b1 = tf.constant(1.)
w2 = tf.constant(2.)
b2 = tf.constant(1.)


with tf.GradientTape(persistent=True) as tape:
	# 非tf.Variable类型的张量需要人为设置记录梯度信息
	tape.watch([w1, b1, w2, b2])
	# 构建2层网络
	y1 = x * w1 + b1	
	y2 = y1 * w2 + b2

# 独立求解出各个导数
dy2_dy1 = tape.gradient(y2, [y1])[0]
dy1_dw1 = tape.gradient(y1, [w1])[0]
dy2_dw1 = tape.gradient(y2, [w1])[0]

# 验证链式法则
print(dy2_dy1 * dy1_dw1)
print(dy2_dw1)