Sparse Tensors and TFRecords

Thursday April 27, 2017

When a matrix, array, or tensor has lots of values that are zero, it can be called sparse. You might want to represent the zeros implicitly with a sparse representation. TensorFlow has support for this, and the support extends to its TFRecords Example format.

Here is a sparse one-dimensional tensor:

[0, 7, 0, 0, 8, 0, 0, 0, 0]

The tensor is sparse, in that it has a lot of zeros, but the representation is dense, in that all those zeros are represented explicitly.

A sparse representation of the same tensor will focus only on the non-zero values.

values = [7, 8]

We have to also remember where those values occur, by their indices:

indices = [1, 5]

The one-dimensional indices form will work with some methods, for this one-dimensional example, but in general indices have multiple dimensions, so it will be more consistent (and work everywhere) to represent indices like this:

indices = [[1], [5]]

With values and indices, we don't have quite enough information yet. How many zeros are there to the right of the last value? We have to represent the dense shape of the tensor.

dense_shape = [9]

These three things together, values, indices, and dense_shape, are a sparse representation of the tensor.

TensorFlow accepts lists of values and NumPy arrays to define dense tensors, and it returns NumPy arrays when dense tensors are evaluated. But what to do with sparse tensors? SciPy has several sparse matrix representations, but not a good match for TensorFlow's general sparse tensor form. So for sparse tensors, instead of reusing an existing Python class, TensorFlow provides tf.SparseTensorValue. These are values that exist outside the TensorFlow graph, so they can be made without a tf.Session, for example.

tf.SparseTensorValue(values=values, indices=indices, dense_shape=dense_shape)
## SparseTensorValue(indices=[[1], [5]], values=[7, 8], dense_shape=[9])

Using tf.SparseTensor puts that in the TensorFlow graph.

tf.SparseTensor(values=values, indices=indices, dense_shape=dense_shape)
## <tensorflow.python.framework.sparse_tensor.SparseTensor at 0x11a4e0c10>

That tf.SparseTensor will be constant, since we specified all the pieces of it, and if you run it in a session, you'll get back the equivalent tf.SparseTensorValue.

TensorFlow has operations specifically for working with sparse tensors, such as tf.sparse_matmul. And you can change a sparse matrix to a dense one with tf.sparse_tensor_to_dense. These operations live in the graph, so they have to be run to see a result.

sparse = tf.SparseTensor(values=values, indices=indices, dense_shape=dense_shape)
dense = tf.sparse_tensor_to_dense(sparse)
session.run(dense)
## array([0, 7, 0, 0, 0, 8, 0, 0, 0], dtype=int32)

Going from dense to sparse seems a little less straightforward at the moment, so let's continue assuming we already have the components of our sparse representation.

Going to more dimensions is quite natural. Here's a two-dimensional tensor with three non-zero values:

[[0, 0, 0, 0, 0, 7],
 [0, 5, 0, 0, 0, 0],
 [0, 0, 0, 0, 9, 0],
 [0, 0, 0, 0, 0, 0]]

This can be represented in sparse form as:

indices = [[0, 5],
           [1, 1],
           [2, 4]]

values = [7, 5, 9]

dense_shape = [4, 6]

tf.SparseTensorValue(values=values, indices=indices, dense_shape=dense_shape)
## SparseTensorValue(indices=[[0, 5], [1, 1], [2, 4]], values=[7, 5, 9], dense_shape=[4, 6])

Now, to represent this in a TFRecords Example requires a little bit of transformation. TFRecords only support lists of integers, floats, and bytestrings. The values are easily represented in one Feature, but to represent the indices, each dimension will need its own Feature in the Example. The dense_shape isn't represented at all; that's left to be specified at parsing.

my_example = tf.train.Example(features=tf.train.Features(feature={
    'index_0': tf.train.Feature(int64_list=tf.train.Int64List(value=[0, 1, 2])),
    'index_1': tf.train.Feature(int64_list=tf.train.Int64List(value=[5, 1, 4])),
    'values': tf.train.Feature(int64_list=tf.train.Int64List(value=[7, 5, 9]))
}))
my_example_str = my_example.SerializeToString()

This TFRecord sparse representation can then be parsed inside the graph as a tf.SparseFeature.

my_example_features = {'sparse': tf.SparseFeature(index_key=['index_0', 'index_1'],
                                                  value_key='values',
                                                  dtype=tf.int64,
                                                  size=[4, 6])}
serialized = tf.placeholder(tf.string)
parsed = tf.parse_single_example(serialized, features=my_example_features)
session.run(parsed, feed_dict={serialized: my_example_str})
## {'sparse': SparseTensorValue(indices=array([[0, 5], [1, 1], [2, 4]]),
##                              values=array([7, 5, 9]),
##                              dense_shape=array([4, 6]))}

Support for multi-dimensional sparse features seems to be new in TensorFlow 1.1, and TensorFlow gives this warning when you use SparseFeature:

WARNING:tensorflow:SparseFeature is a complicated feature config
                   and should only be used after careful consideration
                   of VarLenFeature.

VarLenFeature doesn't support real sparsity or multi-dimensionality though; it only supports "ragged edges" as in the case when one example has three elements and the next has seven, for example.

It is a little awkward to put together a sparse representation for TFRecords, but it does give you a lot of flexibility. To put a point on it, I don't know what you can do with a SequenceExample that you can't do with a regular Example using all of FixedLenFeature, VarLenFeature, and SparseFeature.


I'm working on Building TensorFlow systems from components, a workshop at OSCON 2017.