percept

The goal of percept is to allow users to easily fit neural network models to their data.

Installation

You can install the development version of percept from GitHub with:

# install.packages("devtools")
devtools::install_github("hrryt/percept")

Example

This is a basic example of a feed-forward neural network using the MNIST database:

library(percept)

read_mnist <- function(file) {
  x <- as.matrix(read.csv(file, header = FALSE, colClasses = "integer"))
  dimnames(x) <- NULL
  x
}

# MNIST data formatted to CSV by Joseph Redmon
mnist_train <- read_mnist("https://pjreddie.com/media/files/mnist_train.csv")
mnist_test <- read_mnist("https://pjreddie.com/media/files/mnist_test.csv")

digits <- 0:9
correct_train <- factor(mnist_train[, 1], levels = digits)
expected_train <- to_categorical(correct_train)
input_train <- mnist_train[, -1] / 255

model <- Sequential$new(list(
  Dense$new(20, Sigmoid),
  Dense$new(20, ReLU),
  Dense$new(10, Sigmoid)
))

system.time(
  model$fit(input_train, expected_train, batch_size=100, epochs=2000, learning_rate=1)
)
#> 200 
#> 400 
#> 600 
#> 800 
#> 1000 
#> 1200 
#> 1400 
#> 1600 
#> 1800 
#> 2000
#>    user  system elapsed 
#>  24.557   0.053  24.616

How does the model perform on the test set?

correct_test <- factor(mnist_test[, 1], levels = digits)
expected_test <- to_categorical(correct_test)
input_test <- mnist_test[, -1] / 255

output_test <- model$feed_forward(input_test)
predicted_test <- to_factor(output_test, labels=digits)
mean(predicted_test == correct_test)
#> [1] 0.9134

What does the model look like?

plot(model)

How does each input node tend to contribute to the first hidden layer?

w <- colSums(model$layers[[1]]$weights)
w <- (w / max(abs(w)) + 1) / 2
colors <- rgb(colorRamp(c("red", "white", "blue"))(w), maxColorValue = 255)
raster <- as.raster(matrix(colors, 28, 28))
plot(raster)