基于TensorFlow中自定义梯度的2种方式

#使用修饰器，建立梯度反向传播函数。其中op.input包含输入值、输出值，grad包含上层传来的梯度
@tf.RegisterGradient("QuantizeGrad")
def sign_grad(op, grad):
 input = op.inputs[0] # 取出当前的输入
 cond = (input>=-1)&(input<=1) # 大于1或者小于-1的值的位置
 zeros = tf.zeros_like(grad) # 定义出0矩阵用于掩膜
 return tf.where(cond, grad, zeros) 
 # 将大于1或者小于-1的上一层的梯度置为0
 
#使用with上下文管理器覆盖原始的sign梯度函数
def binary(input):
 x = input
 with tf.get_default_graph().gradient_override_map({"Sign":'QuantizeGrad'}):
 #重写梯度
  x = tf.sign(x)
 return x
 
#使用
x = binary(x)

tf.identity(
 input,
 name=None
)

import tensorflow as tf
x1 = tf.Variable(1)
x2 = tf.Variable(3)
x3 = tf.Variable(6)
@tf.RegisterGradient('CustomGrad')
def CustomGrad(op, grad):
#  tf.Print(grad)
 return -grad
 
g = tf.get_default_graph()
oo = x1+x2
with g.gradient_override_map({"Identity": "CustomGrad"}):
 output = tf.identity(oo)
grad_1 = tf.gradients(output, oo)
with tf.Session() as sess:
 sess.run(tf.global_variables_initializer())
 print(sess.run(grad_1))

AttributeError: 'int' object has no attribute 'name'

t = g(x)
y = t + tf.stop_gradient(f(x) - t)

import tensorflow as tf

x1 = tf.Variable(1)
x2 = tf.Variable(3)
x3 = tf.Variable(6)

f = x1+x2*x3
t = -f

y1 = t + tf.stop_gradient(f-t)
y2 = f

grad_1 = tf.gradients(y1, x1)
grad_2 = tf.gradients(y2, x1)
with tf.Session(config=config) as sess:
 sess.run(tf.global_variables_initializer())

 print(sess.run(grad_1))
 print(sess.run(grad_2))