Question

我希望程序可以

用x复制tf.identity的旧值，
将新值y分配给x

但是tf.identity并没有保留x的旧值。

tf.control_dependencies应该要求在执行步骤2之前执行步骤1。

结果：

version: 1.14.1-dev20190531
x_before_: 0.0
x_ident_: 1.0
x_after_: 1.0

测试：

import numpy as np
import tensorflow as tf

dtype = np.float32
x = tf.get_variable('x', shape=(), dtype=dtype,
                    initializer=tf.zeros_initializer)
y = tf.constant(1, dtype=dtype)

# Copy value of x before assigning new value y to x
x_ident = tf.identity(x)
with tf.control_dependencies([x_ident]):
    assign_op = tf.assign(x, y)

# Run
init_op = x.initializer
with tf.Session() as sess:
    sess.run(init_op)
    x_before_ = sess.run(x)
    x_ident_ = sess.run([x_ident, assign_op])[0]
    x_after_ = sess.run(x)

# Check
print("version:", tf.__version__)
print("x_before_:", x_before_)
print("x_ident_:", x_ident_)
print("x_after_:", x_after_)
assert np.allclose(x_ident_, 0)

Answer 1

我过去曾发现此问题，我认为您是对的，它应该可以按预期工作。也许值得filing an issue。我没有这样的解决方案，但是我的解决方法是进行无用的操作，例如添加零：

import numpy as np
import tensorflow as tf

dtype = np.float32
x = tf.get_variable('x', shape=(), dtype=dtype,
                    initializer=tf.zeros_initializer)
y = tf.constant(1, dtype=dtype)

# Force snapshot of x with zero addition
x_ident = x + 0
with tf.control_dependencies([x_ident]):
    assign_op = tf.assign(x, y)

# Run
init_op = x.initializer
with tf.Session() as sess:
    sess.run(init_op)
    x_before_ = sess.run(x)
    x_ident_ = sess.run([x_ident, assign_op])[0]
    x_after_ = sess.run(x)

# Check
print("version:", tf.__version__)
print("x_before_:", x_before_)
print("x_ident_:", x_ident_)
print("x_after_:", x_after_)
assert np.allclose(x_ident_, 0)

输出：

version: 1.14.0
x_before_: 0.0
x_ident_: 0.0
x_after_: 1.0

tf.identity在赋值之前不会复制值

1 个答案: