CODEWITHSUNDEEP
tensorflow
deekay42
deekay42

Reputation: 636

How to repeat tensor elements variable number of times in tensorflow

I have an input tensor which represents an alternation between item and item quantity: 

[item0, qty0, item1, qty1, ...]

I would like to unfold this tensor to

[[item0*qty0], [item1*qty1], ...]

Example:

[1000, 2, 3000, 5, ...]

[[100,1000], [3000,3000,3000,3000,3000], ...]

This is in tensorflow 1.x btw.

Upvotes: 0

Views: 369

Answers (2)

deekay42
deekay42

Reputation: 636

I started with the answer given by zihaozhihao, but since input has dimensions (?) the length is not known and hence tf.unstack won't work.

However, one of the error messages recommended using map_fn and that seems to work:

x_unpacked = tf.reshape(input, (-1, 2))
tiled = tf.map_fn(lambda x: tf.tile([x[0]], [x[1]]), x_unpacked)

Upvotes: 0

zihaozhihao
zihaozhihao

Reputation: 4475

In tf1.x version.

import tensorflow as tf
inputs = tf.constant([10,2,20,3,30,4])
x_unpacked = tf.unstack(tf.reshape(inputs,(-1,2)))
tmp = []
for t in x_unpacked:
    tmp.append(tf.tile([t[0]], [t[1]]))
ans = tf.concat(tmp, axis=0)

with tf.Session() as sess:
    print(sess.run(ans))
# [10 10 20 20 20 30 30 30 30]

In tf2.x, it can be one line,

tf.concat([tf.tile([x[0]],[x[1]]) for x in tf.reshape(inputs, (-1,2))], axis=0)
# [10 10 20 20 20 30 30 30 30]

Upvotes: 1

Related Questions