library(pixmap)
library(MASS)

#só pra printar
library(lattice)
x <- read.pnm(file = "data_atividade2/subject07.normal.pgm", cellres = 1)
plot(x)
polygon(c(45,145,145,45), c(20,20,200,200), border = "Red") #formato do rosto

teste <- (x@grey[40:200,40:150])
levelplot(rotate(teste))
rotate <- function(x) t(apply(x, 2, rev))


data <- matrix(ncol=20691)
for(i in grep("normal|happy|sad|sleepy|surprised|wink", list.files("data_atividade2/"), value = TRUE)){
  path <- paste("data_atividade2", i, sep="/")
  img <- read.pnm(file=path, cellres=1)
  vetor_rosto <- as.numeric(img@grey[40:210, 40:160])
  data <- rbind(data, vetor_rosto)
}
data <- data[-1,]
data <- as.data.frame(data)

datainv<- data[,order(ncol(data):1)]

data <- as.data.frame(rbind(data,datainv))
data$resp <- as.factor(rep(c("happy", "normal", "sad", "sleepy", "surprised","wink"), 30))

set.seed(188288)
fracTraining   <- 0.80
ssTraining   <- floor(fracTraining   * nrow(data))
indexTraining    <- sort(sample(seq_len(nrow(data)), size=ssTraining))
dataTr   <- data[indexTraining, ]
dataTest <- data[-indexTraining, ]

#pca
pc <- prcomp(dataTr[-20692], scale=TRUE, retx=FALSE)
summary(pc)
##35 componentes > 95% variação
##48 com banco espelhado

cumVar <- cumsum(pc$sdev^2/sum(pc$sdev^2))
nCompIdeal <<- max(2, which.max(cumVar >= 0.95))


dataPcaTr <- as.data.frame(predict(pc, dataTr))[1:nCompIdeal]
dataPcaTr$labels <- dataTr$resp
dataPcaTest <- as.data.frame(predict(pc, dataTest))[1:nCompIdeal]
dataPcaTest$labels <- dataTest$resp


fitLDA <- lda(formula = labels ~ . , data=dataPcaTr)
predLDA_Tr <- predict(fitLDA, dataPcaTr)
(conf <- confusionMatrix(predLDA_Tr$class, dataPcaTr$labels))
predLDA_Te <- predict(fitLDA, dataPcaTest)
(conf <- confusionMatrix(predLDA_Te$class, dataPcaTest$labels))

x <- read.pnm(file = "data_atividade2/subjectmarlin.pgm", cellres = 1)
teste <- as.data.frame(rbind(as.numeric(img@grey[40:210, 40:160])))
testePca <- as.data.frame(predict(pc, teste))[1:nCompIdeal]
predict(fitLDA, testePca)