TensorFlow 2.0 Beta is available Learn more

Load text with tf.data

View on TensorFlow.org View source on GitHub Download notebook

This tutorial provides an example of how to use tf.data.TextLineDataset to load examples from text files. TextLineDataset is designed to create a dataset from a text file, in which each example is a line of text from the original file. This is potentially useful for any text data that is primarily line-based (for example, poetry or error logs).

In this tutorial, we'll use three different English translations of the same work, Homer's Illiad, and train a model to identify the translator given a single line of text.


from __future__ import absolute_import, division, print_function, unicode_literals

  # %tensorflow_version only exists in Colab.
  %tensorflow_version 2.x
except Exception:
import tensorflow as tf

import tensorflow_datasets as tfds
import os

The texts of the three translations are by:

The text files used in this tutorial have undergone some typical preprocessing tasks, mostly removing stuff — document header and footer, line numbers, chapter titles. Download these lightly munged files locally.

DIRECTORY_URL = 'https://storage.googleapis.com/download.tensorflow.org/data/illiad/'
FILE_NAMES = ['cowper.txt', 'derby.txt', 'butler.txt']

for name in FILE_NAMES:
  text_dir = tf.keras.utils.get_file(name, origin=DIRECTORY_URL+name)
parent_dir = os.path.dirname(text_dir)

Downloading data from https://storage.googleapis.com/download.tensorflow.org/data/illiad/cowper.txt
819200/815980 [==============================] - 0s 0us/step
Downloading data from https://storage.googleapis.com/download.tensorflow.org/data/illiad/derby.txt
811008/809730 [==============================] - 0s 0us/step
Downloading data from https://storage.googleapis.com/download.tensorflow.org/data/illiad/butler.txt
811008/807992 [==============================] - 0s 0us/step


Load text into datasets

Iterate through the files, loading each one into its own dataset.

Each example needs to be labeled individually labeled, so use tf.data.Dataset.map to apply a labeler function to each one. This will iterate over every example in the dataset, returning (example, label) pairs.

def labeler(example, index):
  return example, tf.cast(index, tf.int64)  

labeled_data_sets = []

for i, file_name in enumerate(FILE_NAMES):
  lines_dataset = tf.data.TextLineDataset(os.path.join(parent_dir, file_name))
  labeled_dataset = lines_dataset.map(lambda ex: labeler(ex, i))

Combine these labeled datasets into a single dataset, and shuffle it.

TAKE_SIZE = 5000
all_labeled_data = labeled_data_sets[0]
for labeled_dataset in labeled_data_sets[1:]:
  all_labeled_data = all_labeled_data.concatenate(labeled_dataset)
all_labeled_data = all_labeled_data.shuffle(
    BUFFER_SIZE, reshuffle_each_iteration=False)

You can use tf.data.Dataset.take and print to see what the (example, label) pairs look like. The numpy property shows each Tensor's value.

for ex in all_labeled_data.take(5):
(<tf.Tensor: id=74, shape=(), dtype=string, numpy=b'"Neptune, do thou determine for thyself'>, <tf.Tensor: id=75, shape=(), dtype=int64, numpy=1>)
(<tf.Tensor: id=78, shape=(), dtype=string, numpy=b'from you with a great ransom. Fear, O Achilles, the wrath of heaven;'>, <tf.Tensor: id=79, shape=(), dtype=int64, numpy=2>)
(<tf.Tensor: id=82, shape=(), dtype=string, numpy=b'From Brysia, from Augeia, from the rocks'>, <tf.Tensor: id=83, shape=(), dtype=int64, numpy=0>)
(<tf.Tensor: id=86, shape=(), dtype=string, numpy=b'Forth to his son, but with a scream, the child'>, <tf.Tensor: id=87, shape=(), dtype=int64, numpy=0>)
(<tf.Tensor: id=90, shape=(), dtype=string, numpy=b"Forward he mov'd, as when gigantic Mars">, <tf.Tensor: id=91, shape=(), dtype=int64, numpy=1>)

Encode text lines as numbers

Machine learning models work on numbers, not words, so the string values need to be converted into lists of numbers. To do that, map each unique word to a unique integer.

Build vocabulary

First, build a vocabulary by tokenizing the text into a collection of individual unique words. There are a few ways to do this in both TensorFlow and Python. For this tutorial:

  1. Iterate over each example's numpy value.
  2. Use tfds.features.text.Tokenizer to split it into tokens.
  3. Collect these tokens into a Python set, to remove duplicates.
  4. Get the size of the vocabulary for later use.
tokenizer = tfds.features.text.Tokenizer()

vocabulary_set = set()
for text_tensor, _ in all_labeled_data:
  some_tokens = tokenizer.tokenize(text_tensor.numpy())

vocab_size = len(vocabulary_set)

Encode examples

Create an encoder by passing the vocabulary_set to tfds.features.text.TokenTextEncoder. The encoder's encode method takes in a string of text and returns a list of integers.

encoder = tfds.features.text.TokenTextEncoder(vocabulary_set)

You can try this on a single line to see what the output looks like.

example_text = next(iter(all_labeled_data))[0].numpy()
b'"Neptune, do thou determine for thyself'
encoded_example = encoder.encode(example_text)
[4677, 11671, 1161, 9115, 6733, 10246]

Now run the encoder on the dataset by wrapping it in tf.py_function and passing that to the dataset's map method.

def encode(text_tensor, label):
  encoded_text = encoder.encode(text_tensor.numpy())
  return encoded_text, label

def encode_map_fn(text, label):
  return tf.py_function(encode, inp=[text, label], Tout=(tf.int64, tf.int64))

all_encoded_data = all_labeled_data.map(encode_map_fn)

Split the dataset into text and train batches

Use tf.data.Dataset.take and tf.data.Dataset.skip to create a small test dataset and a larger training set.

Before being passed into the model, the datasets need to be batched. Typically, the examples inside of a batch need to be the same size and shape. But, the examples in these datasets are not all the same size — each line of text had a different number of words. So use tf.data.Dataset.padded_batch (instead of batch) to pad the examples to the same size.

train_data = all_encoded_data.skip(TAKE_SIZE).shuffle(BUFFER_SIZE)
train_data = train_data.padded_batch(BATCH_SIZE, padded_shapes=([-1],[]))

test_data = all_encoded_data.take(TAKE_SIZE)
test_data = test_data.padded_batch(BATCH_SIZE, padded_shapes=([-1],[]))

Now, test_data and train_data are not collections of (example, label) pairs, but collections of batches. Each batch is a pair of (many examples, many labels) represented as arrays.

To illustrate:

sample_text, sample_labels = next(iter(test_data))

sample_text[0], sample_labels[0]
WARNING: Logging before flag parsing goes to stderr.
W0813 06:04:24.451589 139653587912448 backprop.py:842] The dtype of the watched tensor must be floating (e.g. tf.float32), got tf.string
W0813 06:04:24.453407 139653587912448 backprop.py:842] The dtype of the watched tensor must be floating (e.g. tf.float32), got tf.int64
W0813 06:04:24.455191 139653587912448 backprop.py:842] The dtype of the watched tensor must be floating (e.g. tf.float32), got tf.string
W0813 06:04:24.455993 139653587912448 backprop.py:842] The dtype of the watched tensor must be floating (e.g. tf.float32), got tf.int64
W0813 06:04:24.457554 139653587912448 backprop.py:842] The dtype of the watched tensor must be floating (e.g. tf.float32), got tf.string

(<tf.Tensor: id=149240, shape=(17,), dtype=int64, numpy=
 array([ 4677, 11671,  1161,  9115,  6733, 10246,     0,     0,     0,
            0,     0,     0,     0,     0,     0,     0,     0])>,
 <tf.Tensor: id=149244, shape=(), dtype=int64, numpy=1>)

Since we have introduced a new token encoding (the zero used for padding), the vocabulary size has increased by one.

vocab_size += 1

Build the model

model = tf.keras.Sequential()

The first layer converts integer representations to dense vector embeddings. See the Word Embeddings tutorial for more details.

model.add(tf.keras.layers.Embedding(vocab_size, 64))

The next layer is a Long Short-Term Memory layer, which lets the model understand words in their context with other words. A bidirectional wrapper on the LSTM helps it to learn about the datapoints in relationship to the datapoints that came before it and after it.


Finally we'll have a series of one or more densely connected layers, with the last one being the output layer. The output layer produces a probability for all the labels. The one with the highest probability is the models prediction of an example's label.

# One or more dense layers.
# Edit the list in the `for` line to experiment with layer sizes.
for units in [64, 64]:
  model.add(tf.keras.layers.Dense(units, activation='relu'))

# Output layer. The first argument is the number of labels.
model.add(tf.keras.layers.Dense(3, activation='softmax'))

Finally, compile the model. For a softmax categorization model, use sparse_categorical_crossentropy as the loss function. You can try other optimizers, but adam is very common.


Train the model

This model running on this data produces decent results (about 83%).

model.fit(train_data, epochs=3, validation_data=test_data)
Epoch 1/3

W0813 06:04:44.479855 139659289282304 deprecation.py:323] From /tmpfs/src/tf_docs_env/lib/python3.5/site-packages/tensorflow/python/ops/math_grad.py:1250: add_dispatch_support.<locals>.wrapper (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version.
Instructions for updating:
Use tf.where in 2.0, which has the same broadcast rule as np.where

697/697 [==============================] - 43s 61ms/step - loss: 0.5220 - accuracy: 0.7456 - val_loss: 0.0000e+00 - val_accuracy: 0.0000e+00
Epoch 2/3
697/697 [==============================] - 34s 48ms/step - loss: 0.2981 - accuracy: 0.8691 - val_loss: 0.3833 - val_accuracy: 0.8264
Epoch 3/3
697/697 [==============================] - 34s 49ms/step - loss: 0.2177 - accuracy: 0.9082 - val_loss: 0.4322 - val_accuracy: 0.8198

<tensorflow.python.keras.callbacks.History at 0x7f0480ae4828>
eval_loss, eval_acc = model.evaluate(test_data)

print('\nEval loss: {:.3f}, Eval accuracy: {:.3f}'.format(eval_loss, eval_acc))
     79/Unknown - 3s 41ms/step - loss: 0.4322 - accuracy: 0.8198
Eval loss: 0.432, Eval accuracy: 0.820