Tensorflow, getting index of non-zero values for every row in a matrix

Question

So I want to get the index of the values that are not zero for every row in my matrix. I've tried with tf.where but the output is not like I expecteded.

My code now is:

b = tf.constant([[1,0,0,0,0],
                 [1,0,1,0,1]],dtype=tf.float32)
zero = tf.constant(0, dtype=tf.float32)
where = tf.not_equal(b, zero)
indices = tf.where(where)

And indices output is:

<tf.Tensor: id=136, shape=(4, 2), dtype=int64, numpy=
array([[0, 0],
       [1, 0],
       [1, 2],
       [1, 4]])>

But I would like the output to be:

[[0],
 [0,2,4]]

Where I have a list with the indexes per row.

Thanks.

Answer 1

That cannot be a proper tensor, since dimensions are not uniform. If you are okay with using a ragged tensor you can do:

import tensorflow as tf

with tf.Graph().as_default(), tf.Session() as sess:
    b = tf.constant([[1, 0, 0, 0, 0],
                     [1, 0, 1, 0, 1]],dtype=tf.float32)
    num_rows = tf.shape(b)[0]
    zero = tf.constant(0, dtype=tf.float32)
    where = tf.not_equal(b, zero)
    indices = tf.where(where)
    s = tf.ragged.segment_ids_to_row_splits(indices[:, 0], num_rows)
    row_start = s[:-1]
    elem_per_row = s[1:] - row_start
    idx = tf.expand_dims(row_start, 1) + tf.ragged.range(elem_per_row)
    result = tf.gather(indices[:, 1], idx)
    print(sess.run(result))
    # <tf.RaggedTensorValue [[0], [0, 2, 4]]>

EDIT: If you don't want to or cannot use ragged tensors, here is an alternative. You can produce a tensor padded with "invalid" values. You can either have for example -1 in those invalid values, or just have a 1D tensor that tells you how many valid values you have on each row:

import tensorflow as tf

with tf.Graph().as_default(), tf.Session() as sess:
    b = tf.constant([[1, 0, 0, 0, 0],
                     [1, 0, 1, 0, 1]],dtype=tf.float32)
    num_rows = tf.shape(b)[0]
    zero = tf.constant(0, dtype=tf.float32)
    where = tf.not_equal(b, zero)
    indices = tf.where(where)
    num_indices = tf.shape(indices)[0]
    elem_per_row = tf.bincount(tf.cast(indices[:, 0], tf.int32), minlength=num_rows)
    row_start = tf.concat([[0], tf.cumsum(elem_per_row[:-1])], axis=0)
    max_elem_per_row = tf.reduce_max(elem_per_row)
    r = tf.range(max_elem_per_row)
    idx = tf.expand_dims(row_start, 1) + r
    idx = tf.minimum(idx, num_indices - 1)
    result = tf.gather(indices[:, 1], idx)
    # Optional: replace invalid elements with -1
    result = tf.where(tf.expand_dims(elem_per_row, 1) > r, result, -tf.ones_like(result))
    print(sess.run(result))
    # [[ 0 -1 -1]
    #  [ 0  2  4]]
    print(sess.run(elem_per_row))
    # [1 3]