Working with preprocessing layers

View on TensorFlow.org Run in Google Colab View source on GitHub Download notebook

Keras preprocessing

The Keras preprocessing layers API allows developers to build Keras-native input processing pipelines. These input processing pipelines can be used as independent preprocessing code in non-Keras workflows, combined directly with Keras models, and exported as part of a Keras SavedModel.

With Keras preprocessing layers, you can build and export models that are truly end-to-end: models that accept raw images or raw structured data as input; models that handle feature normalization or feature value indexing on their own.

Available preprocessing

Text preprocessing

Numerical features preprocessing

Categorical features preprocessing

Image preprocessing

These layers are for standardizing the inputs of an image model.

Image data augmentation

These layers apply random augmentation transforms to a batch of images. They are only active during training.

The adapt() method

Some preprocessing layers have an internal state that can be computed based on a sample of the training data. The list of stateful preprocessing layers is:

  • TextVectorization: holds a mapping between string tokens and integer indices
  • StringLookup and IntegerLookup: hold a mapping between input values and integer indices.
  • Normalization: holds the mean and standard deviation of the features.
  • Discretization: holds information about value bucket boundaries.

Crucially, these layers are non-trainable. Their state is not set during training; it must be set before training, either by initializing them from a precomputed constant, or by "adapting" them on data.

You set the state of a preprocessing layer by exposing it to training data, via the adapt() method:

import numpy as np
import tensorflow as tf
from tensorflow.keras import layers

data = np.array([[0.1, 0.2, 0.3], [0.8, 0.9, 1.0], [1.5, 1.6, 1.7],])
layer = layers.Normalization()
layer.adapt(data)
normalized_data = layer(data)

print("Features mean: %.2f" % (normalized_data.numpy().mean()))
print("Features std: %.2f" % (normalized_data.numpy().std()))

Features mean: -0.00
Features std: 1.00

The adapt() method takes either a Numpy array or a tf.data.Dataset object. In the case of StringLookup and TextVectorization, you can also pass a list of strings:

data = [
    "ξεῖν᾽, ἦ τοι μὲν ὄνειροι ἀμήχανοι ἀκριτόμυθοι",
    "γίγνοντ᾽, οὐδέ τι πάντα τελείεται ἀνθρώποισι.",
    "δοιαὶ γάρ τε πύλαι ἀμενηνῶν εἰσὶν ὀνείρων:",
    "αἱ μὲν γὰρ κεράεσσι τετεύχαται, αἱ δ᾽ ἐλέφαντι:",
    "τῶν οἳ μέν κ᾽ ἔλθωσι διὰ πριστοῦ ἐλέφαντος,",
    "οἵ ῥ᾽ ἐλεφαίρονται, ἔπε᾽ ἀκράαντα φέροντες:",
    "οἱ δὲ διὰ ξεστῶν κεράων ἔλθωσι θύραζε,",
    "οἵ ῥ᾽ ἔτυμα κραίνουσι, βροτῶν ὅτε κέν τις ἴδηται.",
]
layer = layers.TextVectorization()
layer.adapt(data)
vectorized_text = layer(data)
print(vectorized_text)

tf.Tensor(
[[37 12 25  5  9 20 21  0  0]
 [51 34 27 33 29 18  0  0  0]
 [49 52 30 31 19 46 10  0  0]
 [ 7  5 50 43 28  7 47 17  0]
 [24 35 39 40  3  6 32 16  0]
 [ 4  2 15 14 22 23  0  0  0]
 [36 48  6 38 42  3 45  0  0]
 [ 4  2 13 41 53  8 44 26 11]], shape=(8, 9), dtype=int64)

In addition, adaptable layers always expose an option to directly set state via constructor arguments or weight assignment. If the intended state values are known at layer construction time, or are calculated outside of the adapt() call, they can be set without relying on the layer's internal computation. For instance, if external vocabulary files for the TextVectorization, StringLookup, or IntegerLookup layers already exist, those can be loaded directly into the lookup tables by passing a path to the vocabulary file in the layer's constructor arguments.

Here's an example where we instantiate a StringLookup layer with precomputed vocabulary:

vocab = ["a", "b", "c", "d"]
data = tf.constant([["a", "c", "d"], ["d", "z", "b"]])
layer = layers.StringLookup(vocabulary=vocab)
vectorized_data = layer(data)
print(vectorized_data)

tf.Tensor(
[[1 3 4]
 [4 0 2]], shape=(2, 3), dtype=int64)

Preprocessing data before the model or inside the model

There are two ways you could be using preprocessing layers:

Option 1: Make them part of the model, like this:

inputs = keras.Input(shape=input_shape)
x = preprocessing_layer(inputs)
outputs = rest_of_the_model(x)
model = keras.Model(inputs, outputs)

With this option, preprocessing will happen on device, synchronously with the rest of the model execution, meaning that it will benefit from GPU acceleration. If you're training on GPU, this is the best option for the Normalization layer, and for all image preprocessing and data augmentation layers.

Option 2: apply it to your tf.data.Dataset, so as to obtain a dataset that yields batches of preprocessed data, like this:

dataset = dataset.map(lambda x, y: (preprocessing_layer(x), y))

With this option, your preprocessing will happen on CPU, asynchronously, and will be buffered before going into the model. In addition, if you call dataset.prefetch(tf.data.AUTOTUNE) on your dataset, the preprocessing will happen efficiently in parallel with training:

dataset = dataset.map(lambda x, y: (preprocessing_layer(x), y))
dataset = dataset.prefetch(tf.data.AUTOTUNE)
model.fit(dataset, ...)

This is the best option for TextVectorization, and all structured data preprocessing layers. It can also be a good option if you're training on CPU and you use image preprocessing layers.

When running on TPU, you should always place preprocessing layers in the tf.data pipeline (with the exception of Normalization and Rescaling, which run fine on TPU and are commonly used as the first layer is an image model).

Benefits of doing preprocessing inside the model at inference time

Even if you go with option 2, you may later want to export an inference-only end-to-end model that will include the preprocessing layers. The key benefit to doing this is that it makes your model portable and it helps reduce the training/serving skew.

When all data preprocessing is part of the model, other people can load and use your model without having to be aware of how each feature is expected to be encoded & normalized. Your inference model will be able to process raw images or raw structured data, and will not require users of the model to be aware of the details of e.g. the tokenization scheme used for text, the indexing scheme used for categorical features, whether image pixel values are normalized to [-1, +1] or to [0, 1], etc. This is especially powerful if you're exporting your model to another runtime, such as TensorFlow.js: you won't have to reimplement your preprocessing pipeline in JavaScript.

If you initially put your preprocessing layers in your tf.data pipeline, you can export an inference model that packages the preprocessing. Simply instantiate a new model that chains your preprocessing layers and your training model:

inputs = keras.Input(shape=input_shape)
x = preprocessing_layer(inputs)
outputs = training_model(x)
inference_model = keras.Model(inputs, outputs)

Quick recipes

Image data augmentation

Note that image data augmentation layers are only active during training (similarly to the Dropout layer).

from tensorflow import keras
from tensorflow.keras import layers

# Create a data augmentation stage with horizontal flipping, rotations, zooms
data_augmentation = keras.Sequential(
    [
        layers.RandomFlip("horizontal"),
        layers.RandomRotation(0.1),
        layers.RandomZoom(0.1),
    ]
)

# Load some data
(x_train, y_train), _ = keras.datasets.cifar10.load_data()
input_shape = x_train.shape[1:]
classes = 10

# Create a tf.data pipeline of augmented images (and their labels)
train_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_dataset = train_dataset.batch(16).map(lambda x, y: (data_augmentation(x), y))


# Create a model and train it on the augmented image data
inputs = keras.Input(shape=input_shape)
x = layers.Rescaling(1.0 / 255)(inputs)  # Rescale inputs
outputs = keras.applications.ResNet50(  # Add the rest of the model
    weights=None, input_shape=input_shape, classes=classes
)(x)
model = keras.Model(inputs, outputs)
model.compile(optimizer="rmsprop", loss="sparse_categorical_crossentropy")
model.fit(train_dataset, steps_per_epoch=5)

Downloading data from https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
170500096/170498071 [==============================] - 6s 0us/step
170508288/170498071 [==============================] - 6s 0us/step
5/5 [==============================] - 11s 44ms/step - loss: 8.8927
<keras.callbacks.History at 0x7f1c0c3f16d0>

You can see a similar setup in action in the example image classification from scratch.

Normalizing numerical features

# Load some data
(x_train, y_train), _ = keras.datasets.cifar10.load_data()
x_train = x_train.reshape((len(x_train), -1))
input_shape = x_train.shape[1:]
classes = 10

# Create a Normalization layer and set its internal state using the training data
normalizer = layers.Normalization()
normalizer.adapt(x_train)

# Create a model that include the normalization layer
inputs = keras.Input(shape=input_shape)
x = normalizer(inputs)
outputs = layers.Dense(classes, activation="softmax")(x)
model = keras.Model(inputs, outputs)

# Train the model
model.compile(optimizer="adam", loss="sparse_categorical_crossentropy")
model.fit(x_train, y_train)

1563/1563 [==============================] - 3s 2ms/step - loss: 2.1304
<keras.callbacks.History at 0x7f1bc43f40d0>

Encoding string categorical features via one-hot encoding

# Define some toy data
data = tf.constant([["a"], ["b"], ["c"], ["b"], ["c"], ["a"]])

# Use StringLookup to build an index of the feature values and encode output.
lookup = layers.StringLookup(output_mode="one_hot")
lookup.adapt(data)

# Convert new test data (which includes unknown feature values)
test_data = tf.constant([["a"], ["b"], ["c"], ["d"], ["e"], [""]])
encoded_data = lookup(test_data)
print(encoded_data)

tf.Tensor(
[[0. 0. 0. 1.]
 [0. 0. 1. 0.]
 [0. 1. 0. 0.]
 [1. 0. 0. 0.]
 [1. 0. 0. 0.]
 [1. 0. 0. 0.]], shape=(6, 4), dtype=float32)

Note that, here, index 0 is reserved for out-of-vocabulary values (values that were not seen during adapt()).

You can see the StringLookup in action in the Structured data classification from scratch example.

Encoding integer categorical features via one-hot encoding

# Define some toy data
data = tf.constant([[10], [20], [20], [10], [30], [0]])

# Use IntegerLookup to build an index of the feature values and encode output.
lookup = layers.IntegerLookup(output_mode="one_hot")
lookup.adapt(data)

# Convert new test data (which includes unknown feature values)
test_data = tf.constant([[10], [10], [20], [50], [60], [0]])
encoded_data = lookup(test_data)
print(encoded_data)

tf.Tensor(
[[0. 0. 1. 0. 0.]
 [0. 0. 1. 0. 0.]
 [0. 1. 0. 0. 0.]
 [1. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0.]
 [0. 0. 0. 0. 1.]], shape=(6, 5), dtype=float32)

Note that index 0 is reserved for missing values (which you should specify as the value 0), and index 1 is reserved for out-of-vocabulary values (values that were not seen during adapt()). You can configure this by using the mask_token and oov_token constructor arguments of IntegerLookup.

You can see the IntegerLookup in action in the example structured data classification from scratch.

Applying the hashing trick to an integer categorical feature

If you have a categorical feature that can take many different values (on the order of 10e3 or higher), where each value only appears a few times in the data, it becomes impractical and ineffective to index and one-hot encode the feature values. Instead, it can be a good idea to apply the "hashing trick": hash the values to a vector of fixed size. This keeps the size of the feature space manageable, and removes the need for explicit indexing.

# Sample data: 10,000 random integers with values between 0 and 100,000
data = np.random.randint(0, 100000, size=(10000, 1))

# Use the Hashing layer to hash the values to the range [0, 64]
hasher = layers.Hashing(num_bins=64, salt=1337)

# Use the CategoryEncoding layer to multi-hot encode the hashed values
encoder = layers.CategoryEncoding(num_tokens=64, output_mode="multi_hot")
encoded_data = encoder(hasher(data))
print(encoded_data.shape)

(10000, 64)

Encoding text as a sequence of token indices

This is how you should preprocess text to be passed to an Embedding layer.

# Define some text data to adapt the layer
adapt_data = tf.constant(
    [
        "The Brain is wider than the Sky",
        "For put them side by side",
        "The one the other will contain",
        "With ease and You beside",
    ]
)

# Create a TextVectorization layer
text_vectorizer = layers.TextVectorization(output_mode="int")
# Index the vocabulary via `adapt()`
text_vectorizer.adapt(adapt_data)

# Try out the layer
print(
    "Encoded text:\n", text_vectorizer(["The Brain is deeper than the sea"]).numpy(),
)

# Create a simple model
inputs = keras.Input(shape=(None,), dtype="int64")
x = layers.Embedding(input_dim=text_vectorizer.vocabulary_size(), output_dim=16)(inputs)
x = layers.GRU(8)(x)
outputs = layers.Dense(1)(x)
model = keras.Model(inputs, outputs)

# Create a labeled dataset (which includes unknown tokens)
train_dataset = tf.data.Dataset.from_tensor_slices(
    (["The Brain is deeper than the sea", "for if they are held Blue to Blue"], [1, 0])
)

# Preprocess the string inputs, turning them into int sequences
train_dataset = train_dataset.batch(2).map(lambda x, y: (text_vectorizer(x), y))
# Train the model on the int sequences
print("\nTraining model...")
model.compile(optimizer="rmsprop", loss="mse")
model.fit(train_dataset)

# For inference, you can export a model that accepts strings as input
inputs = keras.Input(shape=(1,), dtype="string")
x = text_vectorizer(inputs)
outputs = model(x)
end_to_end_model = keras.Model(inputs, outputs)

# Call the end-to-end model on test data (which includes unknown tokens)
print("\nCalling end-to-end model on test string...")
test_data = tf.constant(["The one the other will absorb"])
test_output = end_to_end_model(test_data)
print("Model output:", test_output)

Encoded text:
 [[ 2 19 14  1  9  2  1]]

Training model...
1/1 [==============================] - 3s 3s/step - loss: 0.4776

Calling end-to-end model on test string...
Model output: tf.Tensor([[0.04233753]], shape=(1, 1), dtype=float32)

You can see the TextVectorization layer in action, combined with an Embedding mode, in the example text classification from scratch.

Note that when training such a model, for best performance, you should always use the TextVectorization layer as part of the input pipeline.

Encoding text as a dense matrix of ngrams with multi-hot encoding

This is how you should preprocess text to be passed to a Dense layer.

# Define some text data to adapt the layer
adapt_data = tf.constant(
    [
        "The Brain is wider than the Sky",
        "For put them side by side",
        "The one the other will contain",
        "With ease and You beside",
    ]
)
# Instantiate TextVectorization with "multi_hot" output_mode
# and ngrams=2 (index all bigrams)
text_vectorizer = layers.TextVectorization(output_mode="multi_hot", ngrams=2)
# Index the bigrams via `adapt()`
text_vectorizer.adapt(adapt_data)

# Try out the layer
print(
    "Encoded text:\n", text_vectorizer(["The Brain is deeper than the sea"]).numpy(),
)

# Create a simple model
inputs = keras.Input(shape=(text_vectorizer.vocabulary_size(),))
outputs = layers.Dense(1)(inputs)
model = keras.Model(inputs, outputs)

# Create a labeled dataset (which includes unknown tokens)
train_dataset = tf.data.Dataset.from_tensor_slices(
    (["The Brain is deeper than the sea", "for if they are held Blue to Blue"], [1, 0])
)

# Preprocess the string inputs, turning them into int sequences
train_dataset = train_dataset.batch(2).map(lambda x, y: (text_vectorizer(x), y))
# Train the model on the int sequences
print("\nTraining model...")
model.compile(optimizer="rmsprop", loss="mse")
model.fit(train_dataset)

# For inference, you can export a model that accepts strings as input
inputs = keras.Input(shape=(1,), dtype="string")
x = text_vectorizer(inputs)
outputs = model(x)
end_to_end_model = keras.Model(inputs, outputs)

# Call the end-to-end model on test data (which includes unknown tokens)
print("\nCalling end-to-end model on test string...")
test_data = tf.constant(["The one the other will absorb"])
test_output = end_to_end_model(test_data)
print("Model output:", test_output)

WARNING:tensorflow:5 out of the last 1567 calls to <function PreprocessingLayer.make_adapt_function.<locals>.adapt_step at 0x7f1b9c5c5290> triggered tf.function retracing. Tracing is expensive and the excessive number of tracings could be due to (1) creating @tf.function repeatedly in a loop, (2) passing tensors with different shapes, (3) passing Python objects instead of tensors. For (1), please define your @tf.function outside of the loop. For (2), @tf.function has experimental_relax_shapes=True option that relaxes argument shapes that can avoid unnecessary retracing. For (3), please refer to https://www.tensorflow.org/guide/function#controlling_retracing and https://www.tensorflow.org/api_docs/python/tf/function for  more details.
Encoded text:
 [[1. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 1. 1. 0. 0. 0. 0. 0.

  0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 1. 1. 0. 0. 0.]]

Training model...
1/1 [==============================] - 0s 231ms/step - loss: 1.0046

Calling end-to-end model on test string...
Model output: tf.Tensor([[-0.54753447]], shape=(1, 1), dtype=float32)

Encoding text as a dense matrix of ngrams with TF-IDF weighting

This is an alternative way of preprocessing text before passing it to a Dense layer.

# Define some text data to adapt the layer
adapt_data = tf.constant(
    [
        "The Brain is wider than the Sky",
        "For put them side by side",
        "The one the other will contain",
        "With ease and You beside",
    ]
)
# Instantiate TextVectorization with "tf-idf" output_mode
# (multi-hot with TF-IDF weighting) and ngrams=2 (index all bigrams)
text_vectorizer = layers.TextVectorization(output_mode="tf-idf", ngrams=2)
# Index the bigrams and learn the TF-IDF weights via `adapt()`

with tf.device("CPU"):
    # A bug that prevents this from running on GPU for now.
    text_vectorizer.adapt(adapt_data)

# Try out the layer
print(
    "Encoded text:\n", text_vectorizer(["The Brain is deeper than the sea"]).numpy(),
)

# Create a simple model
inputs = keras.Input(shape=(text_vectorizer.vocabulary_size(),))
outputs = layers.Dense(1)(inputs)
model = keras.Model(inputs, outputs)

# Create a labeled dataset (which includes unknown tokens)
train_dataset = tf.data.Dataset.from_tensor_slices(
    (["The Brain is deeper than the sea", "for if they are held Blue to Blue"], [1, 0])
)

# Preprocess the string inputs, turning them into int sequences
train_dataset = train_dataset.batch(2).map(lambda x, y: (text_vectorizer(x), y))
# Train the model on the int sequences
print("\nTraining model...")
model.compile(optimizer="rmsprop", loss="mse")
model.fit(train_dataset)

# For inference, you can export a model that accepts strings as input
inputs = keras.Input(shape=(1,), dtype="string")
x = text_vectorizer(inputs)
outputs = model(x)
end_to_end_model = keras.Model(inputs, outputs)

# Call the end-to-end model on test data (which includes unknown tokens)
print("\nCalling end-to-end model on test string...")
test_data = tf.constant(["The one the other will absorb"])
test_output = end_to_end_model(test_data)
print("Model output:", test_output)

WARNING:tensorflow:6 out of the last 1568 calls to <function PreprocessingLayer.make_adapt_function.<locals>.adapt_step at 0x7f1b9f6eae60> triggered tf.function retracing. Tracing is expensive and the excessive number of tracings could be due to (1) creating @tf.function repeatedly in a loop, (2) passing tensors with different shapes, (3) passing Python objects instead of tensors. For (1), please define your @tf.function outside of the loop. For (2), @tf.function has experimental_relax_shapes=True option that relaxes argument shapes that can avoid unnecessary retracing. For (3), please refer to https://www.tensorflow.org/guide/function#controlling_retracing and https://www.tensorflow.org/api_docs/python/tf/function for  more details.
Encoded text:
 [[5.461647  1.6945957 0.        0.        0.        0.        0.

  0.        0.        0.        0.        0.        0.        0.
  0.        0.        1.0986123 1.0986123 1.0986123 0.        0.
  0.        0.        0.        0.        0.        0.        0.
  1.0986123 0.        0.        0.        0.        0.        0.
  0.        1.0986123 1.0986123 0.        0.        0.       ]]

Training model...
1/1 [==============================] - 0s 239ms/step - loss: 4.4868

Calling end-to-end model on test string...
Model output: tf.Tensor([[0.25670475]], shape=(1, 1), dtype=float32)

Important gotchas

Working with lookup layers with very large vocabularies

You may find yourself working with a very large vocabulary in a TextVectorization, a StringLookup layer, or an IntegerLookup layer. Typically, a vocabulary larger than 500MB would be considered "very large".

In such case, for best performance, you should avoid using adapt(). Instead, pre-compute your vocabulary in advance (you could use Apache Beam or TF Transform for this) and store it in a file. Then load the vocabulary into the layer at construction time by passing the filepath as the vocabulary argument.

Using lookup layers on a TPU pod or with ParameterServerStrategy.

There is an outstanding issue that causes performance to degrade when using a TextVectorization, StringLookup, or IntegerLookup layer while training on a TPU pod or on multiple machines via ParameterServerStrategy. This is slated to be fixed in TensorFlow 2.7.