Multiply Tensors with different ranks

Question

I would like to multiply two tensors with different ranks, e.g.

tensor A -> (i x j x k x l)
tensor B -> (l x m)

to get another tensor with this dimensions.

tensor C -> (i x j x k x m) = (i x j x k x l)*(l x m)

In tensorflow I have to reshape all the time to get the output as

(A.reshape(-1,l)*B).reshape(i,j,k,m)

Any other operation that does it directly?

Are you trying to multiple these tensors elementwise, or perform a (batch) matrix multiplication? — mrry, Jun 05 '17 at 14:33
It is a matrix multiplication like this (i x j x k x l)*(l x m) = (i x j x k x m). Possibly, it is a batch matrix multiplication. — RobinHood, Jun 06 '17 at 08:09
`tf.matmul` gives rank error (Tensorflow 1.4). `np.matmul` seems to work as expected. — Souradeep Nanda, Nov 26 '17 at 14:05
use tf.tensordot: https://stackoverflow.com/questions/47969305 — Raphael Schumann, Mar 10 '18 at 03:22

score 3 · Answer 1 · answered Sep 09 '18 at 14:49

You need tf.tensordot. Reshaping before real multiplication is not necessary. If you want to contract tensor A and B along the A's last dimension and the B's first dimension, you can multiply tensor A and tensor B in this way:

tf.tensordot(A, B, axes=[[3], [0]])

Multiply Tensors with different ranks

1 Answers1

Linked