本笔记主要记录tf.GradientTape和tf.gradient的用法

import tensorflow as tf
import numpy as np

tf.__version__

#要计算梯度的所有参数计算过程必须放到gradient tape中
#with tf.GradientTape as tape:
w = tf.constant(1.)
x = tf.constant(2.)

with tf.GradientTape() as tape:
    tape.watch([w])
    y = x * w
#使用gradient()计算w的梯度
grad = tape.gradient(y, [w])
print("Graidient of W:")
print(grad)

#向量求梯度
W = tf.convert_to_tensor([8., 7., 6.])
X = tf.convert_to_tensor([3., 3., 3.])
with tf.GradientTape() as tape:
    tape.watch(X)
    Y = X * W
grads = tape.gradient(Y, X)
print("Gradient of X:")
print(grads)


#计算二阶梯度
w = tf.constant(3.0)
b = tf.constant(6.6)

with tf.GradientTape() as tape1:
    tape1.watch([w])
    with tf.GradientTape() as tape2:
        tape2.watch([w])
        y = 5 * w**2 + b
    grad_dy_dw = tape2.gradient(y, [w])
    print("1st-order gradient [dy/dw]:")
    print(grad_dy_dw)
grad2_dy_dw = tape1.gradient(grad_dy_dw, [w])
print("2nd-order gradient [dy/dw]:")
print(grad2_dy_dw)

运行结果：