TensorFlow broadcasting

问题:

Broadcasting is the process of making arrays with different shapes have compatible shapes for arithmetic operations. In numpy, we can broadcast arrays. Does TensorFlow graph support broadcasting similar to the numpy one?

回答1:

yes it is supported. Open a terminal and try this:

import tensorflow as tf

#define tensors
a=tf.constant([[10,20],[30,40]]) #Dimension 2X2
b=tf.constant([5])
c=tf.constant([2,2])
d=tf.constant([[3],[3]])

sess=tf.Session() #start a session

#Run tensors to generate arrays
mat,scalar,one_d,two_d = sess.run([a,b,c,d])

#broadcast multiplication with scalar
sess.run(tf.multiply(mat,scalar))

#broadcast multiplication with 1_D array (Dimension 1X2)
sess.run(tf.multiply(mat,one_d))

#broadcast multiply 2_d array (Dimension 2X1)
sess.run(tf.multiply(mat,two_d))

sess.close()

回答2:

The short answer is yes.

c.f. Tensorflow Math doc

Note: Elementwise binary operations in TensorFlow follow numpy-style broadcasting.

c.f. tf.add() doc, or tf.multiply() doc, etc.:

NOTE: [the operation] supports broadcasting. More about broadcasting here