Tensorflow: tf.matmul() fails when matrices are sparse, even if a_is_sparse is set. Cryptic error.

[davidbernat]

tf.matmul(A, B, a_is_sparse=True) fails on a = ops.convert_to_tensor(a, name="a") with cryptic error.

Example code:

import tensorflow as tf
import numpy as np

x = tf.sparse_placeholder(tf.float32)
y = tf.Variable(tf.random_uniform([9, 9], minval=0.0, maxval=1.0, dtype=tf.float32))

with tf.Session() as sess:
    sess.run(tf.initialize_all_variables())
    indices = np.array([[3, 2], [4, 5]], dtype=np.int64)
    values = np.array([1.0, 2.0], dtype=np.float32)
    shape = np.array([9, 9], dtype=np.int64)
    _ = sess.run(tf.matmul(x, y), feed_dict={
        x: tf.SparseTensorValue(indices, values, shape)})

Error:
line 12, in
_ = sess.run(tf.matmul(x, y, a_is_sparse=True), feed_dict={
File "/usr/local/lib/python2.7/site-packages/tensorflow/python/ops/math_ops.py", line 1189, in matmul
a = ops.convert_to_tensor(a, name="a")
File "/usr/local/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 620, in convert_to_tensor
ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)
File "/usr/local/lib/python2.7/site-packages/tensorflow/python/ops/constant_op.py", line 179, in _constant_tensor_conversion_function
return constant(v, dtype=dtype, name=name)
File "/usr/local/lib/python2.7/site-packages/tensorflow/python/ops/constant_op.py", line 162, in constant
tensor_util.make_tensor_proto(value, dtype=dtype, shape=shape))
File "/usr/local/lib/python2.7/site-packages/tensorflow/python/framework/tensor_util.py", line 421, in make_tensor_proto
tensor_proto.string_val.extend([compat.as_bytes(x) for x in proto_values])
File "/usr/local/lib/python2.7/site-packages/tensorflow/python/util/compat.py", line 44, in as_bytes
raise TypeError('Expected binary or unicode string, got %r' % bytes_or_text)
TypeError: Expected binary or unicode string, got

Is fixed by explicitly converting from sparse to dense tensor before multiplication:

import tensorflow as tf
import numpy as np

x = tf.sparse_placeholder(tf.float32)
z = tf.sparse_tensor_to_dense(x)
y = tf.Variable(tf.random_uniform([9, 9], minval=0.0, maxval=1.0, dtype=tf.float32))

with tf.Session() as sess:
    sess.run(tf.initialize_all_variables())
    indices = np.array([[3, 2], [4, 5]], dtype=np.int64)
    values = np.array([1.0, 2.0], dtype=np.float32)
    shape = np.array([9, 9], dtype=np.int64)
    _ = sess.run(tf.matmul(z, y), feed_dict={
        x: tf.SparseTensorValue(indices, values, shape)})

[aselle]

@concretevitamin, Is this expected to work, or is this a known limitation of sparse tensors.

[concretevitamin]

Flags a_is_sparse and b_is_sparse do not indicate the operands being SparseTensors. Instead, they are algorithmic hints to invoke a more efficient algorithm on the two dense Tensor operands.

To matmul a SparseTensor and a dense Tensor, you could use tf.sparse_tensor_dense_matmul() instead.

[MarvinTeichmann]

Flags a_is_sparse and b_is_sparse do not indicate the operands being SparseTensors. Instead, they are algorithmic hints to invoke a more efficient algorithm on the two dense Tensor operands.

This should be part of the documentation of matmut. It is very unclear what the flag a_is_sparse does.