tf.floor 的梯度是 None

问题描述

tf.floor 返回无梯度
这意味着 floor 操作之前的权重不会更新，对吗?
但我仍然需要梯度来更新权重
是不是很奇怪?有时我们使用floor、ceil...但它们不能提供渐变或者这是正确的操作
和 tf.cast 一样，梯度返回 none

tf.floor return None gradient
it means the weights before floor operation won't be update, right?
but I still need the gradient to update weights
Isn't it weird? sometimes we use floor, ceil... but they can't deliver gradient or this is right action
as well as tf.cast, the gradient return none

这里有一个问题 #897 讨论梯度，但我不不明白为什么 return None 是好的

Here is an issue #897 to discuss the gradient, but I don't understand why return None is good

numa = tf.constant(5.3)  
res1 = numa * 2   
res2 = tf.floor(numa *2)   
print tf.gradients(res1, numa)
# <type 'list'>: [<tf.Tensor 'gradients/mul_grad/Reshape:0' shape=() dtype=float32>]

print tf.gradients(res2, numa)
# None

我可以直接修改math_grad.py吗?需要其他操作吗?
类似于删除 math_grad.pyc?

Can I modify math_grad.py directly? need other action?
something like delete math_grad.pyc?

@ops.RegisterGradient("Floor")
def _FloorGrad(_, unused_grad):
  # return [None]  # before
  return [unused_grad] # I want to modify

推荐答案

TensorFlow 出于实现原因使用 None 来表示 0.如果返回的梯度是真正的梯度 0，我认为这对您没有帮助，因为您想训练.您可以使用 gradient_override_map 将 Identity 操作的梯度替换为 Floor 如下

TensorFlow uses None to represent 0 for implementation reasons. I don't think it would help you if the returned gradient was true gradient 0, since you want to train. You could use gradient_override_map to substitute gradient of Identity op for Floor as follows

tf.reset_default_graph()
x = tf.Variable(10.)
with tf.get_default_graph().gradient_override_map({"Floor": "Identity"}):
    x2 = tf.floor(x)
loss = tf.square(x2)
opt = tf.train.GradientDescentOptimizer(0.1)
train_op = opt.minimize(loss)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
for i in range(10):
    print(sess.run([loss, train_op]))

这篇关于tf.floor 的梯度是 None的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持！