最近在stackoverflow上看到一个问题，链接是In TensorFlow, what is tf.identity used for?。最高票的答案贴了两段代码，说明了使用tf.identity后，才会得到累加的效果，但并未解释其中原因，这篇文章做一些解释，有不对的地方还请大神指点。
首先，从tensor的定义说起，官网的解释为：A tensor is a generalization of vectors and matrices to potentially higher dimensions. tensor可以是Variable，Constant，Placeholder等等。但是，官网上还有一句话值得注意：Unlike tf.Tensor objects, a tf.Variable exists outside the context of a single session.run call.这个说明了variable的特殊性，这也就是 y = x 失效，而 y = tf.identity(x)有效的原因。同时，对于tf.control_dependencies，官网上有这么句话：control_inputs: A list of Operation or Tensor objects which must be executed or computed before running the operations defined in the context. 继续看下面的代码：
import tensorflow as tf
x = tf.Variable(0.0)
print x
x_plus_1 = tf.assign_add(x, 1)
with tf.control_dependencies([x_plus_1]):
y = x
print y
#z=tf.identity(x,name='x')
init = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init)
for i in range(5):
print(sess.run(y))
运行的结果为：
<tf.Variable 'Variable:0' shape=() dtype=float32_ref>
<tf.Variable 'Variable:0' shape=() dtype=float32_ref>
2018-01-06 17:44:49.511158: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.2 AVX AVX2 FMA
0.0
0.0
0.0
0.0
0.0
此时y的类型为tf.Variable，不受tf.control_dependencies控制。我们把y=x换成y=x+0.0，代码变为：
import tensorflow as tf
x = tf.Variable(0.0)
print x
x_plus_1 = tf.assign_add(x, 1)
with tf.control_dependencies([x_plus_1]):
y = x + 0.0
print y
#z=tf.identity(x,name='x')
init = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init)
for i in range(5):
print(sess.run(y))
再次运行代码，输出为：
<tf.Variable 'Variable:0' shape=() dtype=float32_ref>
Tensor("add:0", shape=(), dtype=float32)
2018-01-06 18:01:28.295217: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.2 AVX AVX2 FMA
1.0
2.0
3.0
4.0
5.0
此时，y的类型变成了tensor类型，受tf.control_dependencies的约束，生成y之前，进行了5次自加操作。最后，换成y = tf.identity(x,name='x')。代码如下：
import tensorflow as tf
x = tf.Variable(0.0)
print x
x_plus_1 = tf.assign_add(x, 1)
with tf.control_dependencies([x_plus_1]):
y = tf.identity(x,name='x')
print y
#z=tf.identity(x,name='x')
init = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init)
for i in range(5):
print(sess.run(y))
运行代码，输出为：
/Users/xuyao/anaconda3/envs/python/bin/python /Users/xuyao/PycharmProjects/Test/test_tf_identity.py
<tf.Variable 'Variable:0' shape=() dtype=float32_ref>
2018-01-06 18:51:58.121838: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.2 AVX AVX2 FMA
Tensor("x:0", shape=(), dtype=float32)
1.0
2.0
3.0
4.0
5.0
tf.identity属于tensorflow中的一个ops，跟x = x + 0.0的性质一样，返回一个tensor，受到tf.control_dependencies的约束，所以生效。
至此，文章开头提到的困惑就可以解答了。
以上属于个人观点。转载请注明，谢谢。
