library(dplyr)
library(tidyr)
library(readr)
#### Reading big data sets into R 


library(ggplot2)
library(gplots)
library(lattice)
library(svdvis)
library(softImpute)
library(ElemStatLearn)
library(rpart.plot)
library(randomForest)
library(ranger)
library(rgl)
#########

library(irlba)
library(bigmemory)
library(biglm)
library(biglars)
# BigR - packages for modeling big data but staying in R


#############Handwritten digits#########
data(zip.train)
Numbers<-as.data.frame(zip.train)
Numbers[,1]<-as.factor(Numbers[,1])
names(Numbers)<-c("number",as.character(seq(1,256)))
#### plot sample images of the 10 digits
par(mfrow=c(3,3))
for (zz in (1:9)) {
  iz<-sample(seq(1,dim(Numbers)[1])[Numbers$number==zz],1)
  mm<-matrix(as.numeric(Numbers[iz,-1]),16,16)
  image(mm[1:16,16:1])
}
#### svd decomposition
svd.obj <- svd(zip.train[,-1])
colnames(svd.obj$v) = paste0("V",1:256)
rownames(svd.obj$v) = paste0("Sample",1:256)
svd.scree(svd.obj, subr=5,
          axis.title.x="Full scree plot", axis.title.y="% Var Explained")
#### 3D plot of the scores
plot3d(svd.obj$u[,1:3],col=zip.train[,1]+1)
legend3d("topright", legend = paste('Type', c(unique(zip.train[,1]))), pch = 5, col=seq(1,10), cex=1, inset=c(0.02))

plot3d(zip.train[,sample(seq(2,257),3)],col=zip.train[,1]+1)
legend3d("topright", legend = paste('Type', c(unique(zip.train[,1]))), pch = 5, col=seq(1,10), cex=1, inset=c(0.02))


####

library(glmnet)

ii<-sample(seq(1,dim(Numbers)[1]),1000)
gg<-glmnet(x=as.matrix(Numbers[ii,-1]),Numbers[ii,1],family="multinomial")
par(mfrow=c(4,3))
plot(gg,xvar="lambda")
###

image(gg$beta[[2]])


####
ggm<-glmnet(x=as.matrix(Numbers[ii,-1]),Numbers[ii,1],family="multinomial",type.multinomial="grouped")
par(mfrow=c(3,3))
plot(ggm,xvar="lambda")
###

image(ggm$beta[[2]])

#####

cv.gg<-cv.glmnet(x=as.matrix(Numbers[ii,-1]),y=Numbers[ii,1],family="multinomial")
par(mfrow=c(1,1))
plot(cv.gg)
names(cv.gg)
# CV error for selecting lambda
par(mfrow=c(1,1))
image(do.call(cbind,coef(cv.gg,s="lambda.1se")))

cc<-coef(cv.gg,s="lambda.1se")
par(mfrow=c(3,3))
for (zz in (1:9)) {
  ccm<-matrix(cc[[zz]][-1],16,16)
  image(ccm[1:16,16:1]) }

pp<-predict(cv.gg,newx=as.matrix(Numbers[-ii,-1]),type="class")
table(pp,Numbers[-ii,1])


#####
cv.ggm<-cv.glmnet(x=as.matrix(Numbers[ii,-1]),y=Numbers[ii,1],family="multinomial",type.multinomial="grouped")
par(mfrow=c(1,1))
plot(cv.ggm)
names(cv.ggm)
# CV error for selecting lambda
par(mfrow=c(1,1))
image(do.call(cbind,coef(cv.ggm,s="lambda.1se")))

cc<-coef(cv.ggm,s="lambda.1se")
par(mfrow=c(3,3))
for (zz in (1:9)) {
ccm<-matrix(cc[[zz]][-1],16,16)
image(ccm[1:16,16:1]) }

pp<-predict(cv.ggm,newx=as.matrix(Numbers[-ii,-1]),type="class")
table(pp,Numbers[-ii,1])

#######

cc<-do.call(cbind,coef(cv.gg,s="lambda.1se"))

bb<-matrix(cc[-1,5],16,16)
image(bb[1:16,16:1])


cc<-do.call(cbind,gg$beta)
cc<-cc[,seq(20,950,by=95)]
bb<-matrix(cc[,5],16,16)
image(bb[1:16,16:1])

#####