R interface to Auto-Keras

Auto-Keras is an open source software library for automated machine learning (AutoML). It is developed by DATA Lab at Texas A&M University and community contributors. The ultimate goal of AutoML is to provide easily accessible deep learning tools to domain experts with limited data science or machine learning background. Auto-Keras provides functions to automatically search for architecture and hyperparameters of deep learning models.

Check out the Auto-Keras blogpost at the RStudio TensorFlow for R blog.

Dependencies

• Auto-Keras requires Python 3.6 .

Installation

AutoKeras is currently only available as a GitHub package.

To install it run the following from an R console:

if (!require("remotes"))
  install.packages("remotes")
remotes::install_github("jcrodriguez1989/autokeras")

Then, use the install_autokeras() function to install TensorFlow:

library("autokeras")
install_autokeras()

Docker

Auto-Keras R package has a configured Docker image.

Steps to run it:

From a bash console:

docker pull jcrodriguez1989/r-autokeras:0.1.0
docker run -it jcrodriguez1989/r-autokeras:0.1.0 /bin/bash

Once inside the Docker image, you can run the example R script:

Rscript cifar10_example.R

Examples

CIFAR-10 dataset

library("autokeras")
library("keras")
# Get CIFAR-10 dataset, but not preprocessing needed
cifar10 <- dataset_cifar10()
c(x_train, y_train) %<-% cifar10$train
c(x_test, y_test) %<-% cifar10$test

# Create an image classifier, and train different models for 5 minutes
clf <- model_image_classifier(verbose=TRUE, augment=FALSE) %>% 
  fit(x_train, y_train, time_limit=5*60)

# Get the best trained model
# If it times out, it fits the model anyways
clf %>% final_fit(x_train, y_train, x_test, y_test, retrain=TRUE, time_limit=60)

# And use it to evaluate, predict
clf %>% evaluate(x_test, y_test)

## [1] 0.368

clf %>% predict(x_test[1:10,,,])

##  [1] 6 8 8 8 6 6 9 4 5 9

# get the Keras model to work with the Keras R library
get_keras_model(clf)

## Model
## ___________________________________________________________________________
## Layer (type)                     Output Shape                  Param #     
## ===========================================================================
## input_1 (InputLayer)             (None, 32, 32, 3)             0           
## ___________________________________________________________________________
## activation_1 (Activation)        (None, 32, 32, 3)             0           
## ___________________________________________________________________________
## batch_normalization_1 (BatchNorm (None, 32, 32, 3)             12          
## ___________________________________________________________________________
## conv2d_1 (Conv2D)                (None, 32, 32, 64)            1792        
## ___________________________________________________________________________
## max_pooling2d_1 (MaxPooling2D)   (None, 16, 16, 64)            0           
## ___________________________________________________________________________
## activation_2 (Activation)        (None, 16, 16, 64)            0           
## ___________________________________________________________________________
## batch_normalization_2 (BatchNorm (None, 16, 16, 64)            256         
## ___________________________________________________________________________
## conv2d_2 (Conv2D)                (None, 16, 16, 64)            36928       
## ___________________________________________________________________________
## max_pooling2d_2 (MaxPooling2D)   (None, 8, 8, 64)              0           
## ___________________________________________________________________________
## activation_3 (Activation)        (None, 8, 8, 64)              0           
## ___________________________________________________________________________
## batch_normalization_3 (BatchNorm (None, 8, 8, 64)              256         
## ___________________________________________________________________________
## conv2d_3 (Conv2D)                (None, 8, 8, 64)              36928       
## ___________________________________________________________________________
## max_pooling2d_3 (MaxPooling2D)   (None, 4, 4, 64)              0           
## ___________________________________________________________________________
## global_average_pooling2d_1 (Glob (None, 64)                    0           
## ___________________________________________________________________________
## dropout_1 (Dropout)              (None, 64)                    0           
## ___________________________________________________________________________
## dense_1 (Dense)                  (None, 64)                    4160        
## ___________________________________________________________________________
## activation_4 (Activation)        (None, 64)                    0           
## ___________________________________________________________________________
## dense_2 (Dense)                  (None, 10)                    650         
## ===========================================================================
## Total params: 80,982
## Trainable params: 80,720
## Non-trainable params: 262
## ___________________________________________________________________________

IMDb dataset

library("autokeras")
library("keras")
# Get IMDb dataset
imdb <- dataset_imdb(num_words = 10000)
c(x_train, y_train) %<-% imdb$train
c(x_test, y_test) %<-% imdb$test
# Auto-Keras procceses each text data point as a character vector,
# i.e., x_train[[1]] "<START> this film was just brilliant casting..",
# so we need to transform the dataset.
word_index <- dataset_imdb_word_index()
word_index <- c("<PAD>", "<START>", "<UNK>", "<UNUSED>",
                 names(word_index)[order(unlist(word_index))])
x_train <- lapply(x_train, function(x)
  paste(word_index[x+1], collapse=" "))
x_test <- lapply(x_test, function(x)
  paste(word_index[x+1], collapse=" "))

# Create text classifier, and train different models for 5 minutes
clf <- model_text_classifier(verbose=TRUE) %>%
  fit(x_train, y_train, time_limit=5*60)

# Get the best trained model
# If it times out, it fits the model anyways
clf %>% final_fit(x_train, y_train, x_test, y_test, retrain=TRUE, time_limit=60)

# And use it to evaluate, predict
clf %>% evaluate(x_test, y_test)

## [1] 0.53

clf %>% predict(x_test[1:10])

##  [1] 0 0 1 1 1 1 1 1 1 1

This line does not work, bug already reported in Auto-Keras python library

# get the Keras model to work with the Keras R library
get_keras_model(clf)