Why would I ever use tf.concat instead of tf.stack?

Question

Is there a good reason to use tf.concat instead of tf.stack? They seem very similar. Is it just to guarantee that the resulting tensor will have the same number of dimensions as the input list of tensors?

Answer 1

Actually, I've misunderstood how tf.stack works. If the axis parameter is within the range of the existing dimensions, a new axis will be inserted at that index.

Example:

import tensorflow as tf  t1 = tf.random_normal([1, 3]) t2 = tf.random_normal([1, 3])  tf.stack([t1, t2], axis=1).shape.as_list() == [1, 2, 3] tf.concat([t1, t2], axis=1).shape.as_list() == [1, 6]