为什么我们需要在T​​ensorFlow中“分配”？

Question

I am going through an example here . I see there the following sequence of operations:

import tensorflow as tf

v1 = tf.get_variable("v1", shape=[3], initializer = tf.zeros_initializer)
inc_v1 = v1.assign(v1+1)
init_op = tf.global_variables_initializer()
s = tf.Session()
s.run(init_op)
s.run(inc_v1)

As a result we get:

array([1., 1., 1.], dtype=float32)

I do not understand the logic behind "assign" operation. In particular I have replaced it by something that looks much simpler to me:

import tensorflow as tf

v1 = tf.get_variable("v1", shape=[3], initializer = tf.zeros_initializer)
inc_v1 = v1 + 1
init_op = tf.global_variables_initializer()
s = tf.Session()
s.run(init_op)
s.run(inc_v1)

And I got exactly the same output. So, why do we need assign?

Answer 1

It takes your original tensor and a new tensor, updates original value of your tensor with a new value and returns the reference of your original tensor. Take a look at the graphs generated on Tensorboard:

The operation assign returns a reference to the original Tensor:

Without assign , just creates another tensor to add a constant value:

If you print the evaluation of the tensor v1 (after running inc_v1 ) it outputs [1. 1. 1.] [1. 1. 1.] as the result of the operation its been reassigned to the original tensor. In the second case, it will remain as [0. 0. 0.] [0. 0. 0.] .

Answer 2

That example is indeed not very illustrative. The important part is that assign saves the given value to the variable within the session, so you can use it later in the next call to run . See here:

import tensorflow as tf

v1 = tf.get_variable("v1", shape=[3], initializer = tf.zeros_initializer)
inc_v1 = v1.assign(v1+1)
init_op = tf.global_variables_initializer()
with tf.Session() as sess:
    sess.run(init_op)
    sess.run(inc_v1)
    print(sess.run(v1))
    # [1. 1. 1.]
    sess.run(inc_v1)
    print(sess.run(v1))
    # [2. 2. 2.]

Note v1 saves the assigned value, so in further calls to run it can be used. Compare now to:

import tensorflow as tf

v1 = tf.get_variable("v1", shape=[3], initializer = tf.zeros_initializer)
inc_v1 = v1+1
init_op = tf.global_variables_initializer()
with tf.Session() as sess:
    sess.run(init_op)
    sess.run(inc_v1)
    print(sess.run(v1))
    # [0. 0. 0.]
    sess.run(inc_v1)
    print(sess.run(v1))
    # [0. 0. 0.]

Here the increment happens within one call to run , but its result is never saved anywhere, and v1 keeps having the value [0. 0. 0.] [0. 0. 0.] .

Variables are important because most things in TensorFlow are done on several steps, eg each batch in a neural network. Each step is a call to run , and it is important that the changes to a model in a step (eg the updates to the weights in a neural network) are saved for the next step - otherwise you would be running in place, never leaving the starting point!

Answer 3

如果要分配不是来自图表的值，则需要指定