Tensorflow unsorted_segment_sum dimension

Question

I'm using the tf.unsorted_segment_sum method of TensorFlow and it works fine when the tensor i give as data have only one line. For example:

tf.unsorted_segment_sum(tf.constant([0.2, 0.1, 0.5, 0.7, 0.8]),
                        tf.constant([0, 0, 1, 2, 2]), 3)

Gives the right result:

array([ 0.3,  0.5 , 1.5 ], dtype=float32)

The question is, if i use a tensor with several lines, how can I get the results for each line? For instance, if I try a tensor with two lines:

tf.unsorted_segment_sum(tf.constant([[0.2, 0.1, 0.5, 0.7, 0.8],
                                     [0.2, 0.2, 0.5, 0.7, 0.8]]),
                        tf.constant([[0, 0, 1, 2, 2],
                                     [0, 0, 1, 2, 2]]), 3)

The result i would expect is:

array([ [ 0.3,  0.5 , 1.5 ], [ 0.4, 0.5, 1.5 ] ], dtype=float32)

But what I get is:

array([ 0.7,  1. ,  3. ], dtype=float32)

I want to know if someone know how to obtain the result for each line without using a for loop?

Thanks in advance

Answer 1

EDIT:

While the solution below may cover some additional strange uses, this problem can be solved much more easily just by transposing the data. It turns out that, even though tf.unsorted_segment_sum does not have an axis parameter, it can work only along one axis, as long as it is the first one. So you can do just as follows:

import tensorflow as tf

with tf.Session() as sess:
    data = tf.constant([[0.2, 0.1, 0.5, 0.7, 0.8],
                        [0.2, 0.2, 0.5, 0.7, 0.8]])
    idx = tf.constant([0, 0, 1, 2, 2])
    result = tf.transpose(tf.unsorted_segment_sum(tf.transpose(data), idx, 3))
    print(sess.run(result))

Output:

[[ 0.30000001  0.5         1.5       ]
 [ 0.40000001  0.5         1.5       ]]

---

ORIGINAL POST:

tf.unsorted_segment_sum does not support working on a single axis. The simplest solution would be to apply the operation to each row and then concatenate them back:

data = tf.constant([[0.2, 0.1, 0.5, 0.7, 0.8],
                    [0.2, 0.2, 0.5, 0.7, 0.8]])
segment_ids = tf.constant([[0, 0, 1, 2, 2],
                           [0, 0, 1, 2, 2]])
num_segments = 3
rows = []
for data_i, ids_i in zip(data, segment_ids):
    rows.append(tf.unsorted_segment_sum(data_i, ids_i))
result = tf.stack(rows, axis=0)

However, this has drawbacks: 1) it only works for statically-shaped tensors (that is, you need to have a fixed number of rows) and 2) it may not be as efficient. The first one could be circumvented using a tf.while_loop, but, it would be complicated, and also it would require you to concatenate the rows one by one, which is very inefficient. Also, you already stated you want to avoid loops.

A better option is to use different ids for each row. For example, you could add to each value in segment_id something like num_segments * row_index, so you guarantee that each row will have its own set of ids:

num_rows = tf.shape(segment_ids)[0]
rows_idx = tf.range(num_rows)
segment_ids_per_row = segment_ids + num_segments * tf.expand_dims(rows_idx, axis=1)

Then you can apply the operation and the reshape to get the tensor that you want:

seg_sums = tf.unsorted_segment_sum(data, segment_ids_per_row,
                                   num_segments * num_rows)
result = tf.reshape(seg_sums, [-1, num_segments])

Output:

array([[ 0.3, 0.5, 1.5 ],
       [ 0.4, 0.5, 1.5 ]], dtype=float32)