####################################
##### Color Perception Network #####
####################################

setwd("PATHTODATA")
library(neuralnet)


##### Generate Data

# set.seed(100)
# 
# dat <- data.frame(R = sample(255,500,replace=T),
#                   G = sample(255,500,replace=T),
#                   B = sample(255,500,replace=T))
# 
# write.table(dat,file="colors.txt",row.names=F,quote=F)
# 
# 

##### Read Data

cols       <- read.table("colors.txt",header=T,nrow=250)
cols$label <- as.character(cols$label)
## re-format

cols[,unique(cols$label)] <- 0

for(i in 1:nrow(cols)){
  cols[i,cols[i,"label"]] <- 1
}


##### Check data

table(cols$label)

### any comments on the training set?



##### Define functions to plot colors

plot.col <- function(dat,n){
  plot(c(1,1),type="n",axes=F,xlab="",ylab="",ylim=c(0,2),xlim=c(0,2))
  
  row.n <- n
  row   <- dat[row.n,]
  points(x=1,y=1,cex=35,pch=19,col=rgb(row["R"],row["G"],row["B"],maxColorValue = 255))
}

plot.col(cols,1)


plot.rgb <- function(R,G,B){
  plot(c(1,1),type="n",axes=F,xlab="",ylab="",ylim=c(0,2),xlim=c(0,2))
  
  points(x=1,y=1,cex=35,pch=19,col=rgb(R,G,B,maxColorValue = 255))
}

plot.rgb(255,0,0)



##### Train a network: Predict labels with RGB values
library(neuralnet)

f1 <- paste(unique(cols$label),"+ ",collapse="")
f1 <- gsub(f1,pattern=" \\+ $",replacement="")

formula <- paste(f1,"~ R + G + B")
formula

## training: TRAIN A NEURAL NETWORK WITH NO HIDDEN LAYER


## plot
plot(nn)
plot(nn,intercept = F)



##### Try to predict some color labels

### random colors
set.seed(100)
testset <- data.frame(R = sample(255,10,replace=T),
              G = sample(255,10,replace=T),
              B = sample(255,10,replace=T))

predicted <- predict(nn,testset)
colnames(predicted) <- nn$model.list$response
predicted

plot.col(testset,3)



### focal colors

testset2 <- data.frame(R = 0, G= 0, B = 0)
testset2[1,] <- c(255,0,0)
testset2[2,] <- c(0,255,0)
testset2[3,] <- c(0,0,255)
testset2[4,] <- c(255,255,0)
testset2[5,] <- c(0,255,255)
testset2[6,] <- c(255,0,255)
testset2[7,] <- c(0,0,0)
testset2[8,] <- c(255,255,255)
testset2[9,] <- c(255,120,0)
testset2[10,] <- c(139,69,19)
testset2[11,] <- c(169,169,169)

predicted2 <- predict(nn,testset2)
colnames(predicted2) <- nn$model.list$response
predicted2

plot.col(testset2,11)




##### Inspect the weights

weights <- round(nn$weights[[1]][[1]],5)
rownames(weights) <- c("bias",nn$model.list$variables)
colnames(weights) <- c(nn$model.list$response)

weights

## alternative:
nn$result.matrix


##################################################
##### Another training with one hidden layer #####


## training: TRAIN A NEURALNET WITH ONE HIDDEN LAYER (3 NODES)
## (note: if the model doesn't converge, adjust the training parameters - which ones?)
## (note2: sometimes, the training just stagnates. this happens. just run it again.)

plot(nn2,intercept=F)


##### model predictions
predicted <- compute(nn2,testset)$net.result
colnames(predicted) <- nn2$model.list$response
predicted

plot.col(testset,8)

##### model predictions
predicted2 <- compute(nn2,testset2)$net.result
colnames(predicted2) <- nn2$model.list$response
predicted2

plot.col(testset2,11)


##### Inspect the weights

## input to hidden layer
weights2a <- round(nn2$weights[[1]][[1]],5)
rownames(weights2a) <- c("bias",nn2$model.list$variables)
colnames(weights2a) <- c("node1","node2","node3")

weights2a

## hidden layer to output
weights2b <- round(nn2$weights[[1]][[2]],5)
rownames(weights2b) <- c("bias","node1","node2","node3")
colnames(weights2b) <- c(nn2$model.list$response)

weights2b



## alternative:
nn2$result.matrix



### Inspect the hidden layer
activations <- compute(nn2,testset2)$neurons[[2]]
activations <- round(activations,5)
activations

plot.col(testset2,8)