Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow, 2nd Edition

Appendix F. Special Data Structures

In this appendix we will take a very quick look at the data structures supported by TensorFlow, beyond regular float or integer tensors. This includes strings, ragged tensors, sparse tensors, tensor arrays, sets, and queues.

Strings

Tensors can hold byte strings, which is useful in particular for natural language processing (see Chapter 16):

>>> tf.constant(b"hello world")
<tf.Tensor: id=149, shape=(), dtype=string, numpy=b'hello world'>

If you try to build a tensor with a Unicode string, TensorFlow automatically encodes it to UTF-8:

>>> tf.constant("café")
<tf.Tensor: id=138, shape=(), dtype=string, numpy=b'caf\xc3\xa9'>

It is also possible to create tensors representing Unicode strings. Just create an array of 32-bit integers, each representing a single Unicode code point:¹

>>> tf.constant([ord(c) for c in "café"])
<tf.Tensor: id=211, shape=(4,), dtype=int32,
            numpy=array([ 99,  97, 102, 233], dtype=int32)>

Note

In tensors of type tf.string, the string length is not part of the tensor’s shape. In other words, strings are considered as atomic values. However, in a Unicode string tensor (i.e., an int32 tensor), the length of the string is part of the tensor’s shape.

The tf.strings package contains several functions to manipulate string tensors, such as length() to count the number of bytes in a byte string (or the number of code points if you set unit="UTF8_CHAR"), unicode_encode() to convert a Unicode string tensor (i.e., int32 tensor) to ...