Does tf.math.equal supports tf.sparse.SparseTensor?

[HongtaoYang]

URL with the issue:

https://www.tensorflow.org/api_docs/python/tf/math/equal

Description of issue (what needs changing):

The doc says tf.math.equal supports sparse tensor as inputs, but it is actually not supported at the moment. See the example below. I haven't checked all math operations that claims to support sparse tensors, maybe there are other similar doc errors like this.

I'm using tf 2.6.0-dev20210601.

Usage example

import tensorflow as tf

a = tf.sparse.SparseTensor(
    indices=[[0, 0], [0, 1], [1, 2]],
    values=[1, 1, 1],
    dense_shape=[2, 3]
)

b = tf.sparse.SparseTensor(
    indices=[[0, 0], [0, 1], [1, 2]],
    values=[1, 1, 1],
    dense_shape=[2, 3]
)

tf.math.equal(a, b)  # raises ValueError

I think making equal operation supports sparse tensor is needed.

[tilakrayal]

@Saduf2019
Was able to reproduce the issue with TF v2.4,v2.5 and TF-nightly. Please find the gist of it here. Thanks!

[Namangarg110]

Hey I'm new to contributing to open source, can I work on this issue?

[HongtaoYang]

Thanks for working on the issue. I'm using this temporary hack as a workaround atm.

import tensorflow as tf
from tensorflow.sparse import SparseTensor


def sparse_tensor_not_equal(a: SparseTensor, b: SparseTensor) -> SparseTensor:
    diff = tf.math.abs(tf.sparse.add(tf.math.negative(a), b))
    diff = tf.cast(diff, bool)

    return diff

def sparse_tensor_equal(a: SparseTensor, b: SparseTensor) -> SparseTensor:
    diff = sparse_tensor_not_equal(a, b)
    is_equal = tf.sparse.map_values(tf.math.logical_not, diff)

    return is_equal


a = tf.sparse.SparseTensor(indices=[[0, 0], [0, 1], [1, 2]], values=[1, 1, 1], dense_shape=[2, 3])
b = tf.sparse.SparseTensor(indices=[[0, 1], [0, 2], [1, 1], [1, 2]], values=[1, 1, 1, 1], dense_shape=[2, 3])
print(sparse_tensor_equal(a, b))

[edloper]

Looking at the code, it looks to me like tf.math.equal and tf.math.not_equal do not, and have never, supported sparse tensors. (And they only support IndexedSlices in so far as IndexedSlices can be converted to tensors, so it seems odd to call out support for IndexedSlices.) It looks like this erroneous documentation was added in e0e1efb.

It would certainly be possible to extend tf.math.equal to support sparse tensors. If this is done, we should be sure to document the semantics -- namely that tf.equal(x, y) iff tf.equal(tf.sparse.to_dense(x), tf.sparse.to_dense(y)) (so two sparse tensors are considered equal even if one has an explict nonzero where another has an implicit nonzero). Also, we should document that x and y are assumed to be in standard lexicographic order (and tf.sparse.reorder needs to be used if they are not).

As for other ops in math_ops.py that claim to support SparseTensors, I didn't see any similar documentation errors. In particular, the following claim to work and do work: tf.math.abs, tf.math.negative, tf.math.sign, tf.cast. Also, my_dense * my_sparse works (using the multiplication operator), though I don't see that fact documented anywhere. (In contrast, tf.math.multiply(my_dense, my_sparse) does not work).

[sushreebarsa]

@HongtaoYang
Is this still an issue ?
Could you please refer to the comment above and let us know if it helps?
Thanks!

[HongtaoYang]

Thanks, the comment helps, thanks @edloper ! I'll close this issue.

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No

[edloper]

Fixed the documentation in 011e57d.