Deep Learning with Keras :: Cheatsheet

Hex logo for tensorflow - A red hexagon with a stylized 'TFR' (denoting TensorFlow for R) in a lighter shade of red, with the 'F' joined to the 'R'.

Download PDF



Translations (PDF)

Intro

Keras is a high-level neural networks API developed with a focus on enabling fast experimentation. It supports multiple back-ends, including TensorFlow, Jax and Torch.

Backends like TensorFlow are lower level mathematical libraries for building deep neural network architectures. The keras3 R package makes it easy to use Keras with any backend in R.

  1. Define: Model, Sequential model, Multi-GPU model
  2. Compile: Optimizer, Loss, Metrics
  3. Fit: Batch size, Epochs, Validation split
  4. Evaluate: Evaluate, Plot
  5. Predict: Classes, Probability

Read more at:
https://keras.posit.co
https://www.manning.com/books/deep-learning-with-r-second-edition

Installation

The keras3 R package uses the Python keras library. You can install all the prerequisites directly from R See ?keras3::install_keras for details and options.

library(keras3)
reticulate::install_python()
install_keras()

This installs the required libraries in virtual environment named ‘r-keras’. It will automatically detect if a GPU is available.

Training an Image Recognizer on MNIST Data

The “Hello, World!” of deep learning

# input layer: use MNIST images
mnist <- dataset_mnist()
x_train <- mnist$train$x;  y_train <- mnist$train$y
x_test <- mnist$test$x;  y_test <- mnist$test$y

# reshape and rescale
x_train <- array_reshape(x_train, c(nrow(x_train), 784))
x_test <- array_reshape(x_test, c(nrow(x_test), 784))
x_train <- x_train / 255;  x_test <- x_test / 255

y_train <- to_categorical(y_train, 10)
y_test <- to_categorical(y_test, 10)

# defining the model and layers
model <-
  keras_model_sequential(input_shape = c(28, 28, 1))
model |>
  layer_conv_2d(filters = 32, kernel_size = c(3, 3),
                activation = "relu") |>
  layer_max_pooling_2d(pool_size = c(2, 2)) |>
  layer_conv_2d(filters = 64, kernel_size = c(3, 3),
                activation = "relu") |>
  layer_max_pooling_2d(pool_size = c(2, 2)) |>
  layer_flatten() |>
  layer_dropout(rate = 0.5) |>
  layer_dense(units = num_classes,
              activation = "softmax")

# View the model summary
summary(model)
plot(model)

# compile (define loss and optimizer)
model |> 
  compile(
  loss = 'categorical_crossentropy',
  optimizer = optimizer_rmsprop(),
  metrics = c('accuracy')
)

# train (fit)
model |> fit(
  x_train, y_train,
  epochs = 30, batch_size = 128,
  validation_split = 0.2
)
model |> evaluate(x_test, y_test)
model |> predict(x_test)

# save the full model
save_model(model, "mnist-classifier.keras")

# deploy for serving inference.
dir.create("serving-mnist-classifier")
export_savedmodel(modek, "serving-mnist-classifier/1")
rsconnect::deployTFModel("serving-mnist-classifier")

Working with Keras models

Define a Model

Functional API: keras_input() and keras_model()

Define a Functional Model with inputs and outputs.

inputs <- keras_input(<input-shape>)
outputs <- inputs |>
  layer_dense() |> layer_...
model <- keras_model(inputs, outputs)

Sequential API: keras_model_sequential()

Define a Sequential Model composed of a linear stack of layers

model <-
  keras_model_sequential(<input-shape>) |>
  layer_dense() |> layer_...

Subclassing API: Model()

Subclass the base Model class

Compile a Model

compile(object, optimizer, loss, metrics = NULL): Configure a Keras model for training.

Fit a Model

fit(object, x = NULL, y = NULL, batch_size = NULL, epochs = 10, verbose = 1, callbacks = NULL, ...): Train a Keras model for a fixed number of epochs (iterations)

Customize training:

  • Provide callbacks to fit():
  • Define a custom Callback().
  • Call train_on_batch() in a custom training loop.
  • Subclass Model() and implement a custom train_step method.
  • Write a fully custom training loop. Update weights with model$optimizer$apply(gradients, weights)

Inspect a Model

  • print(model): Print a summary of a Keras model

  • plot(model, show_shapes = FALSE, show_dtype = FALSE, show_layer_names = FALSE, ...): Plot a Keras model

Evaluate a Model

  • evaluate(object, x = NULL, y = NULL, batch_size = NULL): Evaluate a Keras model.

Predict

  • predict(): Generate predictions from a Keras model.

  • predict_on_batch(): Returns predictions for a single batch of samples.

Save/Load a Model

  • save_model(); load_model(): Save/Load models using the “.keras” file format.

  • save_model_weights(); load_model_weights(): Save/load model weights to/from “.h5” files.

  • save_model_config(); load_model_config(): Save/load model architecture to/from a “.json” file.

Core Layers

  • layer_dense(): Add a densely-connected NN layer to an output.

  • layer_einsum_dense(): Add a dense layer with arbitrary dimensionality.

  • layer_activation(): Apply an activation function to an output.

  • layer_dropout(): Applies Dropout to the input.

  • layer_reshape(): Reshapes an output to a certain shape.

  • layer_permute(): Permute the dimensions of an input according to a given pattern.

  • layer_repeat_vector(): Repeats the input n times.

  • layer_lambda(object, f): Wraps arbitrary expression as a layer.

  • layer_activity_regularization(): Layer that applies an update to the cost function based input activity.

  • layer_masking(): Masks a sequence by using a mask value to skip timesteps.

  • layer_flatten(): Flattens an input.

More layers

Convolutional Layers

  • layer_conv_1d(): 1D, e.g. temporal convolution.

  • layer_conv_2d_transpose(): Transposed 2D (deconvolution).

  • layer_conv_2d() : 2D, e.g. spatial convolution over images.

  • layer_conv_3d_transpose(): Transposed 3D (deconvolution).

  • layer_conv_3d(): 3D, e.g. spatial convolution over volumes.

  • layer_conv_lstm_2d(): Convolutional LSTM.

  • layer_separable_conv_2d(): Depthwise separable 2D.

  • layer_upsampling_1d(); layer_upsampling_2d(); layer_upsampling_3d(): Upsampling layer.

  • layer_zero_padding_1d(); layer_zero_padding_2d(); layer_zero_padding_3d(): Zero-padding layer.

  • layer_cropping_1d(); layer_cropping_2d(); layer_cropping_3d(): Cropping layer.

Pooling Layers

  • layer_max_pooling_1d(); layer_max_pooling_2d(); layer_max_pooling_3d(): Maximum pooling for 1D to 3D.

  • layer_average_pooling_1d(); layer_average_pooling_2d(); layer_average_pooling_3d(): Average pooling for 1D to 3D.

  • layer_global_max_pooling_1d(); layer_global_max_pooling_2d(); layer_global_max_pooling_3d(): Global maximum pooling.

  • layer_global_average_pooling_1d(); layer_global_average_pooling_2d(); layer_global_average_pooling_3d(): Global average pooling.

Preprocessing

Image Preprocessing

Load Images

  • image_dataset_from_directory() Create a TF Dataset from image files in a directory.

  • image_load(), image_from_array(), image_to_array(), image_array_save(): Work with PIL Image instances

Transform Images

Operations that transform image tensors in deterministic ways.

  • op_image_crop()
  • op_image_extract_patches()
  • op_image_pad()
  • op_image_resize()
  • op_image_affine_transform()
  • op_image_map_coordinates()
  • op_image_rgb_to_grayscale()

Resize images without aspect ratio distortion.

  • image_smart_resize():

Image Layers

Builtin image preprocessing layers. Note, any image operation function can also be used as a layer in a Model, or used in layer_lambda().

Image Preprocessing Layers
  • layer_resizing()
  • layer_rescaling()
  • layer_center_crop()
Image Augmentation Layers

Preprocessing layers that randomly augment image inputs during training.

  • layer_random_crop()
  • layer_random_flip()
  • layer_random_translation()
  • layer_random_rotation()
  • layer_random_zoom()
  • layer_random_contrast()
  • layer_random_brightness()

Sequence Preprocesing

  • timeseries_dataset_from_array(): Generate a TF Dataset of sliding windows over a timeseries provided as array.

  • audio_dataset_from_directory(): Generate a TF Dataset from audio files.

  • pad_sequences(): Pad sequences to the same length

Text Preprocessing

  • text_dataset_from_directory(): Generate a TF Dataset from text files in a directory.

  • layer_text_vectorization(), get_vocabulary(), set_vocabulary(): Map text to integer sequences.

Numerical Features Preprocessing

  • layer_normalization(): Normalizes continuous features.

  • layer_discretization(): Buckets continuous features by ranges.

Categorical Features Preprocessing

  • layer_category_encoding(): Encode integer features.

  • layer_hashing(): Hash and bin categorical features.

  • layer_hashed_crossing(): Cross features using the “hashing trick”.

  • layer_string_lookup(): Map strings to (possibly encoded) indices.

  • layer_integer_lookup(): Map integers to (possibly encoded) indices.

Tabular Data

One-stop utility for preprocessing and encoding structured data. Define a feature space from a list of table columns (features).

feature_space <- layer_feature_space(features = list(<features>))

Adapt the feature space to a dataset

adapt(feature_space, dataset)

Use the adapted feature_space preprocessing layer as a layer in a Keras Model, or in the data input pipeline with tfdatasets::dataset_map()

Available features:

  • feature_float()
  • feature_float_rescaled()
  • feature_float_normalized()
  • feature_float_discretized()
  • feature_integer_categorical()
  • feature_integer_hashed()
  • feature_string_categorical()
  • feature_string_hashed()
  • feature_cross()
  • feature_custom()

Pre-trained models

Keras applications are deep learning models that are made available alongside pre-trained weights. These models can be used for prediction, feature extraction, and fine-tuning.

MobileNetV3 Model, pre-trained on ImageNet

  • application_mobilenet_v3_large()
  • application_mobilenet_v3_small()

EfficientNetV2 Model, pre-trained on ImageNet

  • application_efficientnet_v2s()
  • application_efficientnet_v2m()
  • application_efficientnet_v2l()

Inception-ResNet v2 and v3 model, with weights trained on ImageNet

  • application_inception_resnet_v2()
  • application_inception_v3()

VGG16 and VGG19 models

  • application_vgg16()
  • application_vgg19()

ResNet50 model

  • application_resnet50():

NASNet model architecture

  • application_nasnet_large()
  • application_nasnet_mobile()

ImageNet is a large database of images with labels, extensively used for deep learning

Preprocesses a tensor encoding a batch of images for an application, and decodes predictions from an application. - application_preprocess_inputs() - application_decode_predictions()

Callbacks

A callback is a set of functions to be applied at given stages of the training procedure. You can use callbacks to get a view on internal states and statistics of the model during training.

  • allback_early_stopping(): Stop training when a monitored quantity has stopped improving.

  • callback_learning_rate_scheduler(): Learning rate scheduler.

  • callback_tensorboard(): TensorBoard basic visualizations.

CC BY SA Posit Software, PBC • info@posit.co • posit.co

Learn more at keras.posit.co.

Updated: 2024-06.

packageVersion("keras3")
[1] '1.0.0'