MNIST machine learning example, plotting a learning curve.

mnist-learning-curve.R

library(caret)
library(doParallel)
library(reshape2)
library(ggplot2)

# Enable parallel processing.
cl <- makeCluster(detectCores())
registerDoParallel(cl)

# Load the MNIST digit recognition dataset into R
# http://yann.lecun.com/exdb/mnist/
# assume you have all 4 files and gunzip'd them
# creates train$n, train$x, train$y  and test$n, test$x, test$y
# e.g. train$x is a 60000 x 784 matrix, each row is one digit (28x28)
# call:  show_digit(train$x[5,])   to see a digit.
# brendan o'connor - gist.github.com/39760 - anyall.org

load_mnist <- function() {
  load_image_file <- function(filename) {
    ret = list()
    f = file(filename,'rb')
    readBin(f,'integer',n=1,size=4,endian='big')
    ret$n = readBin(f,'integer',n=1,size=4,endian='big')
    nrow = readBin(f,'integer',n=1,size=4,endian='big')
    ncol = readBin(f,'integer',n=1,size=4,endian='big')
    x = readBin(f,'integer',n=ret$n*nrow*ncol,size=1,signed=F)
    ret$x = matrix(x, ncol=nrow*ncol, byrow=T)
    close(f)
    ret
  }
  load_label_file <- function(filename) {
    f = file(filename,'rb')
    readBin(f,'integer',n=1,size=4,endian='big')
    n = readBin(f,'integer',n=1,size=4,endian='big')
    y = readBin(f,'integer',n=n,size=1,signed=F)
    close(f)
    y
  }
  train <<- load_image_file('train-images-idx3-ubyte')
  test <<- load_image_file('t10k-images-idx3-ubyte')
  
  train$y <<- load_label_file('train-labels-idx1-ubyte')
  test$y <<- load_label_file('t10k-labels-idx1-ubyte')  
}

show_digit <- function(arr784, col=gray(12:1/12), ...) {
  image(matrix(arr784, nrow=28)[,28:1], col=col, ...)
}

train <- data.frame()
test <- data.frame()

# Load data.
load_mnist()

# Normalize: X = (X - min) / (max - min) => X = (X - 0) / (255 - 0) => X = X / 255.
train$x <- train$x / 255

# Setup training data with digit and pixel values with 60/40 split for train/cv.
inTrain = data.frame(y=train$y, train$x)
inTrain$y <- as.factor(inTrain$y)
trainIndex = createDataPartition(inTrain$y, p = 0.60,list=FALSE)
training = inTrain[trainIndex,]
cv = inTrain[-trainIndex,]

learningData <- data.frame()

for (n in seq(from=500, to=3000, by=500)) {
  print(paste('Start', n))

  row <- data.frame(TrainingAccuracy = 0, CVAccuracy = 0, Iterations = n)
  
  fit <- train(y ~ ., data = head(training, n), method = 'svmRadial', tuneGrid = data.frame(sigma=0.0107249, C=1))
  results1 <- predict(fit, newdata = head(training, n))
  conf <- confusionMatrix(results1, head(training$y, n))
  
  row$TrainingAccuracy <- conf$overall[[1]]
  
  results2 <- predict(fit, newdata = head(cv, n))
  conf <- confusionMatrix(results2, head(cv$y, n))
  
  row$CVAccuracy <- conf$overall[[1]]
  
  learningData <<- rbind(learningData, row)
  
  print(paste('Done', n))

  # Plot learning curve.
  m <- melt(learningData, id.vars='Iterations')
  m <- cbind(m, Iterations=learningData$Iterations)
  qplot(x=Iterations, y=value, colour=variable, data=m) + geom_line(size=1) + theme_bw()
}

# SVM. 95/94.
#fit <- train(y ~ ., data = head(training, 1000), method = 'svmRadial', tuneGrid = data.frame(sigma=0.0107249, C=1))
#results <- predict(fit, newdata = head(cv, 1000))
#confusionMatrix(results, head(cv$y, 1000))