Conditional assignment of tensor values in TensorFlow

Question

I want to replicate the following numpy code in tensorflow. For example, I want to assign a 0 to all tensor indices that previously had a value of 1.

a = np.array([1, 2, 3, 1])
a[a==1] = 0

# a should be [0, 2, 3, 0]

If I write similar code in tensorflow I get the following error.

TypeError: 'Tensor' object does not support item assignment

The condition in the square brackets should be arbitrary as in a[a<1] = 0.

Is there a way to realize this "conditional assignment" (for lack of a better name) in tensorflow?

Answer 1

Comparison operators such as greater than are available within TensorFlow API.

However, there is nothing equivalent to the concise NumPy syntax when it comes to manipulating the tensors directly. You have to make use of individual comparison, where and assign operators to perform the same action.

Equivalent code to your NumPy example is this:

import tensorflow as tf

a = tf.Variable( [1,2,3,1] )    
start_op = tf.global_variables_initializer()    
comparison = tf.equal( a, tf.constant( 1 ) )    
conditional_assignment_op = a.assign( tf.where (comparison, tf.zeros_like(a), a) )

with tf.Session() as session:
    # Equivalent to: a = np.array( [1, 2, 3, 1] )
    session.run( start_op )
    print( a.eval() )    
    # Equivalent to: a[a==1] = 0
    session.run( conditional_assignment_op )
    print( a.eval() )

# Output is:
# [1 2 3 1]
# [0 2 3 0]

The print statements are of course optional, they are just there to demonstrate the code is performing correctly.

Answer 2

I'm also just starting to use tensorflow Maybe some one will fill my approach more intuitive

import tensorflow as tf

conditionVal = 1
init_a = tf.constant([1, 2, 3, 1], dtype=tf.int32, name='init_a')
a = tf.Variable(init_a, dtype=tf.int32, name='a')
target = tf.fill(a.get_shape(), conditionVal, name='target')

init = tf.initialize_all_variables()
condition = tf.not_equal(a, target)
defaultValues = tf.zeros(a.get_shape(), dtype=a.dtype)
calculate = tf.select(condition, a, defaultValues)

with tf.Session() as session:
    session.run(init)
    session.run(calculate)
    print(calculate.eval())

main trouble is that it is difficult to implement "custom logic". if you could not explain your logic within linear math terms you need to write "custom op" library for tensorflow (more details here)