CODEWITHSUNDEEP
pythontensorflow
MrCartoonology
MrCartoonology

Reputation: 2067

Matrix multiplication with tf.sparse_matmul fails with SparseTensor

Why does this not work:

pl_input = tf.sparse_placeholder('float32',shape=[None,30])
W = tf.Variable(tf.random_normal(shape=[30,1]), dtype='float32')
layer1a = tf.sparse_matmul(pl_input, weights, a_is_sparse=True, b_is_sparse=False)

The error message is

TypeError: Failed to convert object of type <class 'tensorflow.python.framework.sparse_tensor.SparseTensor'> to Tensor. Contents: SparseTensor(indices=Tensor("Placeholder_11:0", shape=(?, ?), dtype=int64), values=Tensor("Placeholder_10:0", shape=(?,), dtype=float32), dense_shape=Tensor("Placeholder_9:0", shape=(?,), dtype=int64)). Consider casting elements to a supported type.

I'm hoping to create a SparseTensorValue that I retrieve batches from, then feed a batch into the pl_input.

Upvotes: 1

Views: 1120

Answers (1)

P-Gn
P-Gn

Reputation: 24641

TL;DR

Use tf.sparse_tensor_dense_matmul in place of tf.sparse_matmul; look at the documentation for an alternative using tf.nn.embedding_lookup_sparse.

About sparse matrices and SparseTensors

The problem is not specific to sparse_placeholder, but due to a confusion in tensorflow's terminology.

You have sparse matrices. And then you have SparseTensor. Both are related but different concept.

  • A SparseTensor is a structure that indexes its values and can represent sparse matrices or tensors efficiently.
  • A sparse matrix is a matrix filled mostly with 0. In tensorflow's documentation, it often does not refer to a SparseTensor but to a plain old Tensor filled mostly with 0s.

It is therefore important to look at the expected type of a function's argument to figure out.

So for example, in the documentation of tf.matmul, operands need to be plain Tensors and not SparseTensors, independently of the value of the xxx_is_sparse flags, which explains your error. When these flags are True, what tf.sparse_matmul actually expects really is a (dense) Tensor. In other words, these flags serve some optimization purposes and not input type constraints. (Those optimizations seem to be useful only for rather larger matrices by the way).

Upvotes: 3

Related Questions