xval-noise-models: Fit the models.

xval-noise-models-p2.r

# Compute the SNR.
# Returns "high" values when the model variance is negative to 
# hint the optimizer to get out of there. 
snr <- function(mean, var) {
  negatives <- which(var < 0)
  result <- 20*log10(mean/sqrt(abs(var)))
  result[negatives] <- result[negatives]*8
  result;
}

# Fit the `Optim-SNR` model
fit.optim.snr <- function(sel.channel, param.init) {

  training.pics <- read.csv('train-pics.csv', stringsAsFactors = FALSE, row.names=NULL)
  training.pics <- as.vector(training.pics[,1])
  data.df <- subset(data.frame(vvm.all$var.df), 
                   channel == sel.channel & 
                   pict %in% training.pics)

  real.snr <- snr(data.df$mean, data.df$var)

  target <- function(par) {
    b0 <- par[1]
    b1 <- par[2]
    b2 <- par[3]

    predicted.var <- (b2*data.df$mean + b1)*data.df$mean + b0
    pred.snr <- snr(data.df$mean, predicted.var)
    squared.avg <- mean((real.snr - pred.snr)^2)

    squared.avg;
  }
  
  solve <- optim(par=param.init, fn=target, method='Nelder-Mead',
                 control = list(
                 trace=2,
                 maxit=1000))
}

# Fit the regular 'VVM' model
fit.w.robust <- function(sel.channel) {
  
  training.pics <- read.csv('train-pics.csv', stringsAsFactors = FALSE, row.names=NULL)
  training.pics <- as.vector(training.pics[,1])
  data.df <- subset(data.frame(vvm.all$var.df), 
                   channel == sel.channel & 
                   pict %in% training.pics)
  library(robustbase)
  fit <- lmrob(var ~ mean + I(mean^2), weight=1/mean^2, data=data.df)
  fit;
}