We’re excited to announce that the keras package deal is now out there on CRAN. The package deal gives an R interface to Keras, a high-level neural networks API developed with a give attention to enabling quick experimentation. Keras has the next key options:

• Permits the identical code to run on CPU or on GPU, seamlessly.

• Consumer-friendly API which makes it simple to rapidly prototype deep studying fashions.

• Constructed-in assist for convolutional networks (for laptop imaginative and prescient), recurrent networks (for sequence processing), and any mixture of each.

• Helps arbitrary community architectures: multi-input or multi-output fashions, layer sharing, mannequin sharing, and so forth. Which means that Keras is suitable for constructing basically any deep studying mannequin, from a reminiscence community to a neural Turing machine.

• Is able to operating on high of a number of back-ends together with TensorFlow, CNTK, or Theano.

In case you are already acquainted with Keras and wish to leap proper in, try https://tensorflow.rstudio.com/keras which has the whole lot it’s essential get began together with over 20 full examples to be taught from.

To be taught a bit extra about Keras and why we’re so excited to announce the Keras interface for R, learn on!

Keras and Deep Studying

Curiosity in deep studying has been accelerating quickly over the previous few years, and several other deep studying frameworks have emerged over the identical time-frame. Of all of the out there frameworks, Keras has stood out for its productiveness, flexibility and user-friendly API. On the identical time, TensorFlow has emerged as a next-generation machine studying platform that’s each extraordinarily versatile and well-suited to manufacturing deployment.

Not surprisingly, Keras and TensorFlow have of late been pulling away from different deep studying frameworks:

Google net search curiosity round deep studying frameworks over time. For those who bear in mind This fall 2015 and Q1-2 2016 as complicated, you were not alone. pic.twitter.com/1f1VQVGr8n

— François Chollet (@fchollet) June 3, 2017

The excellent news about Keras and TensorFlow is that you simply don’t want to decide on between them! The default backend for Keras is TensorFlow and Keras may be built-in seamlessly with TensorFlow workflows. There’s additionally a pure-TensorFlow implementation of Keras with deeper integration on the roadmap for later this yr.

Keras and TensorFlow are the cutting-edge in deep studying instruments and with the keras package deal now you can entry each with a fluent R interface.

Getting Began

Set up

To start, set up the keras R package deal from CRAN as follows:

The Keras R interface makes use of the TensorFlow backend engine by default. To put in each the core Keras library in addition to the TensorFlow backend use the install_keras() perform:

This can give you default CPU-based installations of Keras and TensorFlow. If you’d like a extra custom-made set up, e.g. if you wish to make the most of NVIDIA GPUs, see the documentation for install_keras().

MNIST Instance

We will be taught the fundamentals of Keras by strolling by a easy instance: recognizing handwritten digits from the MNIST dataset. MNIST consists of 28 x 28 grayscale photos of handwritten digits like these:

The dataset additionally consists of labels for every picture, telling us which digit it’s. For instance, the labels for the above photos are 5, 0, 4, and 1.

Making ready the Information

The MNIST dataset is included with Keras and may be accessed utilizing the dataset_mnist() perform. Right here we load the dataset then create variables for our take a look at and coaching knowledge:

library(keras)
mnist <- dataset_mnist()
x_train <- mnist$practice$x
y_train <- mnist$practice$y
x_test <- mnist$take a look at$x
y_test <- mnist$take a look at$y

The x knowledge is a three-D array (photos,width,peak) of grayscale values. To organize the information for coaching we convert the three-D arrays into matrices by reshaping width and peak right into a single dimension (28×28 photos are flattened into size 784 vectors). Then, we convert the grayscale values from integers ranging between 0 to 255 into floating level values ranging between 0 and 1:

# reshape
dim(x_train) <- c(nrow(x_train), 784)
dim(x_test) <- c(nrow(x_test), 784)
# rescale
x_train <- x_train / 255
x_test <- x_test / 255

The y knowledge is an integer vector with values starting from 0 to 9. To organize this knowledge for coaching we one-hot encode the vectors into binary class matrices utilizing the Keras to_categorical() perform:

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

Defining the Mannequin

The core knowledge construction of Keras is a mannequin, a option to arrange layers. The best kind of mannequin is the sequential mannequin, a linear stack of layers.

We start by making a sequential mannequin after which including layers utilizing the pipe (%>%) operator:

mannequin <- keras_model_sequential() 
mannequin %>% 
  layer_dense(items = 256, activation = "relu", input_shape = c(784)) %>% 
  layer_dropout(fee = 0.4) %>% 
  layer_dense(items = 128, activation = "relu") %>%
  layer_dropout(fee = 0.3) %>%
  layer_dense(items = 10, activation = "softmax")

The input_shape argument to the primary layer specifies the form of the enter knowledge (a size 784 numeric vector representing a grayscale picture). The ultimate layer outputs a size 10 numeric vector (possibilities for every digit) utilizing a softmax activation perform.

Use the abstract() perform to print the main points of the mannequin:

Mannequin
________________________________________________________________________________
Layer (kind)                        Output Form                    Param #     
================================================================================
dense_1 (Dense)                     (None, 256)                     200960      
________________________________________________________________________________
dropout_1 (Dropout)                 (None, 256)                     0           
________________________________________________________________________________
dense_2 (Dense)                     (None, 128)                     32896       
________________________________________________________________________________
dropout_2 (Dropout)                 (None, 128)                     0           
________________________________________________________________________________
dense_3 (Dense)                     (None, 10)                      1290        
================================================================================
Whole params: 235,146
Trainable params: 235,146
Non-trainable params: 0
________________________________________________________________________________

Subsequent, compile the mannequin with acceptable loss perform, optimizer, and metrics:

mannequin %>% compile(
  loss = "categorical_crossentropy",
  optimizer = optimizer_rmsprop(),
  metrics = c("accuracy")
)

Coaching and Analysis

Use the match() perform to coach the mannequin for 30 epochs utilizing batches of 128 photos:

historical past <- mannequin %>% match(
  x_train, y_train, 
  epochs = 30, batch_size = 128, 
  validation_split = 0.2
)

The historical past object returned by match() consists of loss and accuracy metrics which we are able to plot:

Consider the mannequin’s efficiency on the take a look at knowledge:

mannequin %>% consider(x_test, y_test,verbose = 0)

$loss
[1] 0.1149

$acc
[1] 0.9807

Generate predictions on new knowledge:

mannequin %>% predict_classes(x_test)

  [1] 7 2 1 0 4 1 4 9 5 9 0 6 9 0 1 5 9 7 3 4 9 6 6 5 4 0 7 4 0 1 3 1 3 4 7 2 7 1 2
 [40] 1 1 7 4 2 3 5 1 2 4 4 6 3 5 5 6 0 4 1 9 5 7 8 9 3 7 4 6 4 3 0 7 0 2 9 1 7 3 2
 [79] 9 7 7 6 2 7 8 4 7 3 6 1 3 6 9 3 1 4 1 7 6 9
 [ reached getOption("max.print") -- omitted 9900 entries ]

Keras gives a vocabulary for constructing deep studying fashions that’s easy, elegant, and intuitive. Constructing a query answering system, a picture classification mannequin, a neural Turing machine, or every other mannequin is simply as easy.

The Information to the Sequential Mannequin article describes the fundamentals of Keras sequential fashions in additional depth.

Examples

Over 20 full examples can be found (particular due to [@dfalbel](https://github.com/dfalbel) for his work on these!). The examples cowl picture classification, textual content technology with stacked LSTMs, question-answering with reminiscence networks, switch studying, variational encoding, and extra.

Studying Extra

After you’ve develop into acquainted with the fundamentals, these articles are subsequent step:

• Information to the Sequential Mannequin. The sequential mannequin is a linear stack of layers and is the API most customers ought to begin with.

• Information to the Practical API. The Keras practical API is the best way to go for outlining advanced fashions, corresponding to multi-output fashions, directed acyclic graphs, or fashions with shared layers.

• Coaching Visualization. There are all kinds of instruments out there for visualizing coaching. These embody plotting of coaching metrics, actual time show of metrics inside the RStudio IDE, and integration with the TensorBoard visualization instrument included with TensorFlow.

• Utilizing Pre-Educated Fashions. Keras consists of a variety of deep studying fashions (Xception, VGG16, VGG19, ResNet50, InceptionVV3, and MobileNet) which are made out there alongside pre-trained weights. These fashions can be utilized for prediction, function extraction, and fine-tuning.

• Often Requested Questions. Covers many further matters together with streaming coaching knowledge, saving fashions, coaching on GPUs, and extra.

Keras gives a productive, extremely versatile framework for creating deep studying fashions. We will’t wait to see what the R group will do with these instruments!