During execution, we may need to maintain the state of the ops. We can do so by using tf.Variable(). Let's check out an example declaration of tf.Variable(). This line will create a variable called counter and initialize it to scalar value 0:

state = tf.Variable(0, name="counter") 

Here are the ops to assign a value to the variable:

one = tf.constant(1) 
update = tf.assign(state, one) 

If you are working with variables, we have to initialize them all at once using the following function:

init_op = tf.initialize_all_variables() 

After initialization, we have to run the graph for putting this into effect. We can run the previous ops using the following code:

sess = tf.Session() 
sess.run(init_op) 
print(sess.run(state)) 
sess.run(update)