MNIST_baseline.R

library(XBART)
library(xgboost)

path <- "~/Dropbox/MNIST/"
D <- read.csv(paste(path, "mnist_train.csv", sep = ""), header = FALSE)
y <- D[, 1]
D <- D[, -1]
#
Dtest <- read.csv(paste(path, "mnist_test.csv", sep = ""), header = FALSE)
ytest <- Dtest[, 1]
Dtest <- Dtest[, -1]
pred <- matrix(0, 10000, 10)

X_train <- D
X_test <- Dtest
p <- ncol(X_train)

# for (i in 1:p){
#   X_train[, i] = X_train[, i] + 0.01*rnorm(length(y))
#   X_test[, i] = X_test[, i] + 0.01*rnorm(length(ytest))
# }


X.train <- X_train
X.test <- X_test

v <- 0
for (h in 1:p) {
    breaks <- unique(as.numeric(quantile(c(X_train[, h], X_test[, h]), seq(0, 1, length.out = 4))))
    # breaks = seq(min(c(X_train[,h],X_test[,h])),max(c(X_train[,h],X_test[,h])),length.out = 25)
    breaks <- c(-Inf, breaks, +Inf)
    # print(breaks)
    if (length(breaks) > 3) {
        v <- v + 1
        X.train[, v] <- cut(X_train[, h], breaks = breaks, include.lowest = TRUE, labels = FALSE)
        X.test[, v] <- cut(X_test[, h], breaks = breaks, include.lowest = TRUE, labels = FALSE)
    }
}

# print(v)
X_train <- X.train[, 1:v]
X_test <- X.test[, 1:v]
p <- v

# X_train[,1] = X_train[,1] + 0.01*rnorm(length(y))

X_train <- as.matrix(X_train)
X_test <- as.matrix(X_test)

t <- proc.time()

xgb.basic.mod1 <- xgboost(
    data = X_train, label = y,
    num_class = 10,
    max_depth = 5,
    subsample = 0.80,
    nrounds = 50,
    early_stopping_rounds = 2,
    eta = 0.9,
    params = list(objective = "multi:softprob")
)

t <- proc.time() - t
xgb.basic.pred <- predict(xgb.basic.mod1, X_test)
xgb.basic.pred <- matrix(xgb.basic.pred, ncol = 10, byrow = TRUE)
pred.xgb <- max.col(xgb.basic.pred) - 1

spr <- split(xgb.basic.pred, row(xgb.basic.pred))
logloss.xgb <- sum(unlist(mapply(function(x, y) -log(x[y]), spr, ytest, SIMPLIFY = TRUE)))

cat("running time ", t[3], " seconds \n")

cat("xgboost error rate ", mean(pred.xgb != ytest), "\n")

cat(paste("xgboost logloss : ", round(logloss.xgb, 3)), "\n")


for (i in 0:9) {
    cat(
        "xgboost error rate in ", i, ": ", round(mean(pred.xgb[ytest == i] != i), 4),
        " misclassified as ", tail(names(sort(table(pred.xgb[ytest == i]))), 2)[1], "\n "
    )
}