nnet applied to noisy circle image

hinomaru_NN.R

# resulting image is
# http://dl.dropbox.com/u/15259519/images/hinomaru_NN.png

sr <- function() source("hinomaru_NN.R")

W <<- 300
H <<- 200
R <<- 60
N <<- 6000
E <<- 20

rgbstr <- function(c) {
  sprintf("#%02x%02x%02x",c[1],c[2],c[3])
}

beni.iro <<- rgbstr(c(215, 19, 69))

center <- c(W/2,H/2)
in.circle <- function(x,y) colSums((rbind(x,y)-center)^2)<=R^2

prepare <- function() {
  x <- runif(N,max=W)
  y <- runif(N,max=H)
  z <- as.integer(in.circle(x,y))
  x1 <- (x+rnorm(N,0,E)) %% W;
  y1 <- (y+rnorm(N,0,E)) %% H;
  z1 <- as.integer(in.circle(x1,y1))
  df <- data.frame(x=x1,y=y1,observe=z,answer=z1)
  return(df)
}
  
myplot2d <- function(df,main1='') {
  plot(df[df$observe==0,][1:2],
       xlim=c(0,W),ylim=c(0,H),xlab='',ylab='',axes=F)
  par(new=T)
  plot(df[df$observe==1,][1:2],
	main=main1,
       xlim=c(0,W),ylim=c(0,H),col=beni.iro)
}

myplot3d <- function(df) {
  library(scatterplot3d)
  scatterplot3d(df$x,df$y,df$observe, angle=70,
                pch=20,highlight.3d=T,col.axis="blue",col.grid="lightblue")
}

cross <- function(x,y) {
  cbind(rep(x),c(t(matrix(rep(y,length(x)),length(y)))))
}

learn.nnet <- function(df,size1=12,maxi1=1000) {
  library(nnet)

  xyo <- data.frame(df[1:2],z=factor(df$observe==1))
  nn <- nnet(z~.,size=size1,maxi=maxi1,data=xyo)

  res <- predict(nn,df[1:2],type="class")
  print(table(df$answer,res))

  grid <- cross((0:(W/5))*5,(0:(H/5))*5)
  xy <- data.frame(x=grid[,1],y=grid[,2])
  res <- predict(nn,xy,type="class")

  xy$observe <- as.integer(res==T)
  print(summary(xy))
  return(xy)
}

main <- function() {
  layout(t(matrix(1:6,2,3)))

  df <- prepare()
  myplot2d(df,sprintf("学習用データ(ノイズの標準偏差=%d)",E))
  myplot3d(df)

  df1 <- learn.nnet(df,size1=5)
  myplot2d(df1,"3層ニューラルネット(中間ノード数=5)")
  myplot3d(df1)

  df1 <- learn.nnet(df,size1=7)
  myplot2d(df1,"3層ニューラルネット(中間ノード数=7)")
  myplot3d(df1)

  layout(1)
}

main()