How to print the value of a Tensor object in TensorFlow?

Question

I have been using the introductory example of matrix multiplication in TensorFlow.

matrix1 = tf.constant([[3., 3.]])
matrix2 = tf.constant([[2.],[2.]])
product = tf.matmul(matrix1, matrix2)

When I print the product, it is displaying it as a Tensor object:

<tensorflow.python.framework.ops.Tensor object at 0x10470fcd0>

But how do I know the value of product?

The following doesn't help:

print product
Tensor("MatMul:0", shape=TensorShape([Dimension(1), Dimension(1)]), dtype=float32)

I know that graphs run on Sessions, but isn't there any way I can check the output of a Tensor object without running the graph in a session?

Answer

The easiest[A] way to evaluate the actual value of a Tensor object is to pass it to the Session.run() method, or call Tensor.eval() when you have a default session (i.e. in a with tf.Session(): block, or see below). In general[B], you cannot print the value of a tensor without running some code in a session.

If you are experimenting with the programming model, and want an easy way to evaluate tensors, the tf.InteractiveSession lets you open a session at the start of your program, and then use that session for all Tensor.eval() (and Operation.run()) calls. This can be easier in an interactive setting, such as the shell or an IPython notebook, when it's tedious to pass around a Session object everywhere. For example, the following works in a Jupyter notebook:

with tf.Session() as sess:  print(product.eval()) 

This might seem silly for such a small expression, but one of the key ideas in Tensorflow 1.x is deferred execution : it's very cheap to build a large and complex expression, and when you want to evaluate it, the back-end (to which you connect with a Session) is able to schedule its execution more efficiently (e.g. executing independent parts in parallel and using GPUs).

[A]: To print the value of a tensor without returning it to your Python program, you can use the tf.print() operator, as Andrzej suggests in another answer. According to the official documentation:

To make sure the operator runs, users need to pass the produced op to tf.compat.v1.Session's run method, or to use the op as a control dependency for executed ops by specifying with tf.compat.v1.control_dependencies([print_op]), which is printed to standard output.

Also note that:

In Jupyter notebooks and colabs, tf.print prints to the notebook cell outputs. It will not write to the notebook kernel's console logs.

[B]: You might be able to use the tf.get_static_value() function to get the constant value of the given tensor if its value is efficiently calculable.