张量流张量的命名

Question

Below is a code snippet which attempts to name the result of a tensor operation so I can access it after the network has been saved and restored

   def createForward(self):

    # forward propogation

    Z      = tf.add(tf.matmul(self.W,self.prevLayer.A),self.b)
    self.Z = tf.nn.dropout(Z,self.keepProb,name = self.name+'_Z')
    print(self.name+'_Z',self.Z)

When self.name is 'output' I am expecting the print statement to print out

output_Z Tensor("output_Z:0", shape=(3, ?), dtype=float32)

What I actually get is

output_Z Tensor("output_Z/mul:0", shape=(3, ?), dtype=float32)

Could somebody explain what is happening.

Thanks

Answer 1

I think what you are not familiar with is the naming of the operation in TensorFlow. Let's first see this:

In [2]: import tensorflow as tf      
In [4]: w = tf.Variable([[1,2,3], [4,5,6], [7,8,9], [3,1,5], [4,1,7]], dtype=tf.float32) 
In [6]: z = tf.nn.dropout(w, 0.4, name="output_Z")
In [7]: z.op.name
Out[7]: u'output_Z/mul' 
In [8]: z.name
Out[8]: u'output_Z/mul:0'

As you can see the name of z differs from the name of the operation z but all of them have the operation name mul appended.

To get what you expected you can do like this:

In [12]: Z = tf.identity(z, name="output_Z")
In [13]: Z.op.name
Out[13]: u'output_Z'
In [14]: Z.name
Out[14]: u'output_Z:0'

---

Reference :

Names of ops in tf.variable_scope()

We discussed how tf.variable_scope governs the names of variables. But how does it influence the names of other ops in the scope? It is natural that ops created inside a variable scope should also share that name. For this reason, when we do with tf.variable_scope("name"), this implicitly opens a tf.name_scope("name"). For example:

with tf.variable_scope("foo"):
    x = 1.0 + tf.get_variable("v", [1])
assert x.op.name == "foo/add"