Is it allowed to assign a value to a variable before it enter a computational graph?

Question

I define a simple computational graph involving a variable. When I change a value of the variable it has an expected influence on the output of the computational graph (so, everything works fine, as expected):

s = tf.Session()

x = tf.placeholder(tf.float32)
c = tf.Variable([1.0, 1.0, 1.0], tf.float32)

y = x + c

c = tf.assign(c, [3.0, 3.0, 3.0])   
s.run(c)
print 'Y1:', s.run(y, {x : [10.0, 20.0, 30.0]})

c = tf.assign(c, [2.0, 2.0, 2.0])
s.run(c)
print 'Y2:',  s.run(y, {x : [10.0, 20.0, 30.0]})

When I call this code I get:

Y1: [ 13.  23.  33.]
Y2: [ 12.  22.  32.]

So, the values after the Y1 and Y2 are different, as expected, because they are calculated with different values of c.

The problems start if I assign a value to the variable c before I define how it is involved into calculation of y. In this case I cannot assign a new value of c.

s = tf.Session()

x = tf.placeholder(tf.float32)
c = tf.Variable([1.0, 1.0, 1.0], tf.float32)

c = tf.assign(c, [4.0, 4.0, 4.0])   # this is the line that causes problems
y = x + c

c = tf.assign(c, [3.0, 3.0, 3.0])   
s.run(c)
print 'Y1:', s.run(y, {x : [10.0, 20.0, 30.0]})

c = tf.assign(c, [2.0, 2.0, 2.0])
s.run(c)
print 'Y2:',  s.run(y, {x : [10.0, 20.0, 30.0]})

As the output I get:

Y1: [ 14.  24.  34.]
Y2: [ 14.  24.  34.]

As you can see, each time I calculate y, I get results involving the old values of c. Why is that?

aseipel · Accepted Answer · 2017-10-04 14:01:15Z

2

With TensorFlow, always keep in mind that you're building a computation graph. In your first code snippet, you basically define y = tf.placeholder(tf.float32) + tf.Variable([1.0, 1.0, 1.0], tf.float32). In your second example, you define y = tf.placeholder(tf.float32) + tf.assign(tf.Variable([1.0, 1.0, 1.0], tf.float32), [4.0, 4.0, 4.0]).

So, no matter which value you assign to c, the computation graph contains the assign operation and will always assign [4.0, 4.0, 4.0] to it before computing the sum.

answered Oct 4, 2017 at 14:01

aseipel

7387 silver badges16 bronze badges

Sign up to request clarification or add additional context in comments.

Comments

Jie.Zhou · Accepted Answer · 2017-10-04 13:58:49Z

1

I think this is because that you define the the add operation y = x + c right after c = tf.assign(c, [4.0, 4.0, 4.0]), so each time you run y out, c = tf.assign(c, [4.0, 4.0, 4.0]) this op will always be excuted and although other assign operations will also be excuted but don't affect the final result.

answered Oct 4, 2017 at 13:58

Jie.Zhou

1,30811 silver badges17 bronze badges

Collectives™ on Stack Overflow

Is it allowed to assign a value to a variable before it enter a computational graph?

2 Answers 2

Comments

Comments

Your Answer

Linked

Hot Network Questions

Collectives™ on Stack Overflow

2 Answers 2

Comments

Comments

Your Answer

Sign up or log in

Post as a guest

Linked

Related