CODEWITHSUNDEEP
pythontensorflow
User
User

Reputation: 826

Clarification of the tf.Session() scope TensorFlow

I have a python program in which I have defined a network and as usual, I'm training it inside a function where I have

with tf.Session() as sess:
...
for epoch in xrange(num_epochs):
...
for n in xrange(num_batches):
       _, c = sess.run([optimizer, loss], feed_dict={....

In the loss function, I have to work a lot to get the loss and in particular, I must take the maximum inside a tensor and use it to do stuff. Here an example

values = tf.constant([0, 1, 2, 0, 2], dtype=tf.float32)
max_values = tf.reduce_max(values) # Tensor...
...

in the max_values line if I use the debug it says that it is a tensor and not a value, so if I change my code in this way passing to the function the session created in the previous piece of code

values = tf.constant([0, 1, 2, 0, 2], dtype=tf.float32)
max_values = sess.run(tf.reduce_max(values)) # 2.0
...

it works. But this loss function is already in a scope of a session so my question is why the result is a tensor and not a number? Is there a way to get the value without passing the session to the loss function?

Upvotes: 2

Views: 638

Answers (2)

saket
saket

Reputation: 378

Use Tensor.eval() function to convert a Tensor to its value. In the following example, you can get the value of max_values Tensor.

def loss():
  values = tf.constant([0, 1, 2, 0, 2], dtype=tf.float32)
  max_values = tf.reduce_max(values)
  print (max_values.eval())

with tf.Session() as sess:
  loss()

If you called loss() outside Session scope, you got an error.

Also you can use Eager execution mode. https://www.tensorflow.org/tutorials/eager/eager_basics

Upvotes: 1

Guilherme Uzeda
Guilherme Uzeda

Reputation: 236

According to the documentation:

TensorFlow uses the tf.Session class to represent a connection between the client program---typically a Python program, although a similar interface is available in other languages---and the C++ runtime.

It means that when you do values = tf.constant([0, 1, 2, 0, 2], dtype=tf.float32) you are just inserting a node to your tensorflow Graph! Since Python is the high level API for a low level C++ runtime, you actually need a Session to evaluate your Python code in this low level runtime.

That is why every time you need to calculate or evaluate a Tensorflow variable/method/constant/etc you need to run it in a Session with tf.Session().run(yournode)

I hope it helped

Upvotes: 2

Related Questions