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.
layer_text_vectorization
: turns raw strings into an encoded
representation that can be read by an Embedding
layer or Dense
layer.layer_normalization
: performs feature-wise normalization of
input features.layer_discretization
: turns continuous numerical features
into integer categorical features.layer_category_encoding
: turns integer categorical features
into one-hot, multi-hot, or count dense representations.layer_hashing
: performs categorical feature hashing, also known as
the "hashing trick".layer_string_lookup
: turns string categorical values into an encoded
representation that can be read by an Embedding
layer or Dense
layer.layer_integer_lookup
: turns integer categorical values into an
encoded representation that can be read by an Embedding
layer or Dense
layer.These layers are for standardizing the inputs of an image model.
layer_resizing
: resizes a batch of images to a target size.layer_rescaling
: rescales and offsets the values of a batch of
images (e.g. go from inputs in the [0, 255]
range to inputs in the [0, 1]
range.layer_center_crop
: returns a center crop of a batch of images.These layers apply random augmentation transforms to a batch of images. They are only active during training.
layer_random_clip
layer_random_flip
layer_random_translation
layer_random_rotation
layer_random_zoom
layer_random_contrast
adapt()
methodSome 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 indicesStringLookup
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:
library(keras3) data <- rbind( c(0.1, 0.2, 0.3), c(0.8, 0.9, 1.0), c(1.5, 1.6, 1.7) ) layer <- layer_normalization() layer %>% adapt(data) normalized_data <- layer(data) op_mean(normalized_data) op_std(normalized_data)
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 <- c( "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 <- layer_text_vectorization() layer %>% adapt(data) vectorized_text <- layer(data) vectorized_text
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 you instantiate a StringLookup
layer with precomputed vocabulary:
vocab <- c("a", "b", "c", "d") data <- rbind(c("a", "c", "d"), c("d", "z", "b")) layer <- layer_string_lookup(vocabulary=vocab) vectorized_data <- layer(data) vectorized_data
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 a 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 %>% dataset_map(function(x, y) list(preprocessing_layer(x), y))
With this option, your preprocessing will happen on a 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 %>% dataset_map(function(x, y) list(preprocessing_layer(x), y)) %>% 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 a CPU and you use image preprocessing
layers.
Note that the TextVectorization
layer can only be executed on a CPU, as it is mostly a
dictionary lookup operation. Therefore, if you are training your model on a GPU or a TPU,
you should put the TextVectorization
layer in the tf.data
pipeline to get the best performance.
When running on a TPU, you should always place preprocessing layers in the tf$data
pipeline
(with the exception of Normalization
and Rescaling
, which run fine on a TPU and are commonly
used as the first layer in an image model).
Even if you go with option 2, you may later want to export an inference-only end-to-end model that will include the preprocessing layers. The key benefit to doing this is that it makes your model portable and it helps reduce the training/serving skew.
When all data preprocessing is part of the model, other people can load and use your
model without having to be aware of how each feature is expected to be encoded &
normalized. Your inference model will be able to process raw images or raw structured
data, and will not require users of the model to be aware of the details of e.g. the
tokenization scheme used for text, the indexing scheme used for categorical features,
whether image pixel values are normalized to [-1, +1]
or to [0, 1]
, etc. This is
especially powerful if you're exporting
your model to another runtime, such as TensorFlow.js: you won't have to
reimplement your preprocessing pipeline in JavaScript.
If you initially put your preprocessing layers in your tf.data
pipeline,
you can export an inference model that packages the preprocessing.
Simply instantiate a new model that chains
your preprocessing layers and your training model:
inputs <- keras_input(shape=input_shape) x <- preprocessing_layer(inputs) outputs <- training_model(x) inference_model <- keras_model(inputs, outputs)
Preprocessing layers are compatible with the tf$distribute API for running training across multiple machines.
In general, preprocessing layers should be placed inside a tf.distribute.Strategy.scope()
and called either inside or before the model as discussed above.
with (strategy$scope(), { inputs <- keras_input(shape=input_shape) preprocessing_layer = layer_hashing(10) dense_layer = tf.keras.layers.Dense(16) })
For more details, refer to the Data preprocessing section of the Distributed input tutorial.
Note that image data augmentation layers are only active during training (similarly to
the Dropout
layer).
# Create a data augmentation stage with horizontal flipping, rotations, zooms data_augmentation <- keras_model_sequential() %>% layer_random_flip("horizontal") %>% layer_random_rotation(0.1) %>% layer_random_zoom(0.1) # Load some data c(c(x_train, y_train), .) %<-% dataset_cifar10() input_shape <- dim(x_train)[-1] classes <- 10 # Create a tf.data pipeline of augmented images (and their labels) train_dataset <- tfdatasets::tensor_slices_dataset(list(x_train, y_train)) %>% tfdatasets::dataset_batch(16) %>% tfdatasets::dataset_map(function(x, y) list(data_augmentation(x), y)) # Create a model and train it on the augmented image data inputs <- keras_input(shape=input_shape) x <- layer_rescaling(inputs, 1.0 / 255)# Rescale inputs outputs <- application_resnet50( # Add the rest of the model weights=NULL, 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)
You can see a similar setup in action in the example image classification from scratch.
# Load some data c(c(x_train, y_train), .) %<-% dataset_cifar10() x_train <- array_reshape(x_train, c(dim(x_train)[1], -1)) input_shape <- dim(x_train)[-1] classes <- 10 # Create a Normalization layer and set its internal state using the training data normalizer <- layer_normalization() normalizer %>% adapt(x_train) # Create a model that include the normalization layer inputs <- keras_input(shape=input_shape) x <- normalizer(inputs) outputs <- layer_dense(x, classes, activation="softmax") model <- keras_model(inputs, outputs) # Train the model model %>% compile(optimizer="adam", loss="sparse_categorical_crossentropy") model %>% fit(x_train, y_train, epochs = 1)
# Define some toy data data <- rbind("a", "b", "c", "b", "c", "a") # Use StringLookup to build an index of the feature values and encode output. lookup <- layer_string_lookup(output_mode="one_hot") lookup %>% adapt(data) # Convert new test data (which includes unknown feature values) test_data <- rbind("a", "b", "c", "b", "c", "") encoded_data <- lookup(test_data) encoded_data
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.
# Define some toy data data <- rbind(10, 20, 20, 10, 30, 0) # Use IntegerLookup to build an index of the feature values and encode output. lookup <- layer_integer_lookup(output_mode="one_hot") lookup %>% adapt(data) # Convert new test data (which includes unknown feature values) test_data <- rbind(10, 10, 20, 50, 60, 0) encoded_data <- lookup(test_data) encoded_data
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.
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 <- random_integer(0, 100000, shape=shape(10000, 1)) # Use the Hashing layer to hash the values to the range [0, 64] hasher <- layer_hashing(num_bins=64, salt=1337) # Use the CategoryEncoding layer to multi-hot encode the hashed values encoder <- layer_category_encoding(num_tokens=64, output_mode="multi_hot") encoded_data <- encoder(hasher(data)) encoded_data$shape
This is how you should preprocess text to be passed to an Embedding
layer.
# Define some text data to adapt the layer adapt_data = c( "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 <- layer_text_vectorization(output_mode="int") # Index the vocabulary via `adapt()` text_vectorizer %>% adapt(adapt_data) # Try out the layer text_vectorizer(rbind("The Brain is deeper than the sea")) # Create a simple model inputs <- keras_input(shape=shape(NULL), dtype="int64") outputs <- inputs %>% layer_embedding(input_dim=text_vectorizer$vocabulary_size(), output_dim=16) %>% layer_gru(units=8) %>% layer_dense(units=1) model <- keras_model(inputs, outputs) # Create a labeled dataset (which includes unknown tokens) train_dataset <- tfdatasets::tensor_slices_dataset(list( rbind("The Brain is deeper than the sea", "for if they are held Blue to Blue"), c(1, 0) )) # Preprocess the string inputs, turning them into int sequences train_dataset <- train_dataset %>% tfdatasets::dataset_batch(2) %>% tfdatasets::dataset_map(function(x, y) list(text_vectorizer(x), y)) # Train the model on the int sequences model %>% compile(optimizer="rmsprop", loss="mse") model %>% fit(train_dataset) # For inference, you can export a model that accepts strings as input inputs <- keras_input(shape = 1, dtype="string") x <- text_vectorizer(inputs) outputs <- model(x) end_to_end_model <- keras_model(inputs, outputs) # Call the end-to-end model on test data (which includes unknown tokens) test_data <- rbind("The one the other will absorb") test_output <- end_to_end_model(test_data) test_output
You can see the TextVectorization
layer in action, combined with an Embedding
mode,
in the example
text classification from scratch.
Note that when training such a model, for best performance, you should always
use the TextVectorization
layer as part of the input pipeline.
This is how you should preprocess text to be passed to a Dense
layer.
# Define some text data to adapt the layer adapt_data <- rbind( "The Brain is wider than the Sky", "For put them side by side", "The one the other will contain", "With ease and You beside" ) # Instantiate TextVectorization with "multi_hot" output_mode # and ngrams=2 (index all bigrams) text_vectorizer <- layer_text_vectorization(output_mode="multi_hot", ngrams=2) # Index the bigrams via `adapt()` text_vectorizer %>% adapt(adapt_data) # Try out the layer text_vectorizer(rbind("The Brain is deeper than the sea")) # Create a simple model inputs <- keras_input(shape = text_vectorizer$vocabulary_size()) outputs <- layer_dense(inputs, 1) model <- keras_model(inputs, outputs) # Create a labeled dataset (which includes unknown tokens) train_dataset <- tfdatasets::tensor_slices_dataset(list( rbind("The Brain is deeper than the sea", "for if they are held Blue to Blue"), c(1, 0) )) # Preprocess the string inputs, turning them into int sequences train_dataset <- train_dataset %>% tfdatasets::dataset_batch(2) %>% tfdatasets::dataset_map(function(x,y) list(text_vectorizer(x), y)) # Train the model on the int sequences model %>% compile(optimizer="rmsprop", loss="mse") model %>% fit(train_dataset, epochs = 2) # 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) test_data <- rbind("The one the other will absorb") test_output <- end_to_end_model(test_data) test_output
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 <- rbind( "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 = layer_text_vectorization(output_mode="tf-idf", ngrams=2) # Index the bigrams and learn the TF-IDF weights via `adapt()` text_vectorizer %>% adapt(adapt_data) # Try out the layer text_vectorizer(rbind("The Brain is deeper than the sea")) # Create a simple model inputs <- keras_input(shape = text_vectorizer$vocabulary_size()) outputs <- layer_dense(inputs, 1) model <- keras_model(inputs, outputs) # Create a labeled dataset (which includes unknown tokens) train_dataset <- tfdatasets::tensor_slices_dataset(list( rbind("The Brain is deeper than the sea", "for if they are held Blue to Blue"), c(1, 0) )) # Preprocess the string inputs, turning them into int sequences train_dataset <- train_dataset %>% tfdatasets::dataset_batch(2) %>% tfdatasets::dataset_map(function(x,y) list(text_vectorizer(x), y)) # Train the model on the int sequences 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) test_data <- rbind("The one the other will absorb") test_output <- end_to_end_model(test_data) test_output
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 a 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 file path as the vocabulary
argument.
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.
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.