CODEWITHSUNDEEP
pythontensorflow
Shiania White
Shiania White

Reputation: 455

Multiplying along an arbitrary axis?

If I have the output of a layer with shape [batch_size, height, width, depth] and another tensor of shape [depth], how can I multiply the first tensor by the second such that each slice along the depth direction is multiplied by the corresponding value in the second tensor. That is, if the second tensor is [4, 5, 6], then the multiplication is:

tensor1[:, :, :, 0] * 4
tensor1[:, :, :, 1] * 5
tensor1[:, :, :, 2] * 6

Also, is there a name for this kind of multiplication that I didn't know to search for? Thanks!

Upvotes: 8

Views: 6652

Answers (3)

RobR
RobR

Reputation: 2190

To answer your second question, "is there a name for this kind of multiplication that I didn't know to search for?", it is an element-wise multiplication with broadcasting. Broadcasting refers to operations that implicitly replicate elements of a tensor to make it compatible with a second tensor used in the element-wise operation. Many Tensorflow operations use the same broadcasting methods used by Numpy, which are further described here

Upvotes: 0

rvinas
rvinas

Reputation: 11895

This is straightforward. Just multiply both tensors. For example:

import tensorflow as tf

tensor = tf.Variable(tf.ones([2, 2, 2, 3]))
depth = tf.constant([4, 5, 6], dtype=tf.float32)
result = tensor * depth

sess = tf.Session()
sess.run(tf.initialize_all_variables())
print(sess.run(result))

Upvotes: 7

Dmytro Danevskyi
Dmytro Danevskyi

Reputation: 3159

I've come up with the following: 

a = tf.placeholder(tf.float32, shape = [5, 2])
b = tf.placeholder(tf.float32, shape = 2)

c = tf.concat(1, [tf.mul(x, y) for x, y in zip(tf.split(0, 2, b), tf.split(1, 2, a))])

sess = tf.Session()
print sess.run(c, feed_dict = {a: np.ones([5, 2]), b: [5, 6]})

Looks a little bit odd, but it seems to work fine for me. I've used 2d tensor, but you definetely could extend this to your case.

Upvotes: 0

Related Questions