R Deep Learning Cookbook

Corrupting data to train

1. The next critical function needed to set up a denoising autoencoder is data corruption:

# Add noise using masking or salt & pepper noise methodadd_noise<-function(data, frac=0.10, corr_type=c("masking", "saltPepper", "none")){  if(length(corr_type)>1) corr_type<-corr_type[1]     # Assign a copy of data  data_noise = data    # Evaluate chaining parameters for autoencoder  nROW<-nrow(data)  nCOL<-ncol(data)  nMask<-floor(frac*nCOL)    if(corr_type=="masking"){    for( i in 1:nROW){      maskCol<-sample(nCOL, nMask)      data_noise[i,maskCol,,]<-0    }  } else if(corr_type=="saltPepper"){    minval<-min(data[,,1,])    maxval<-max(data[,,1,])    for( i in 1:nROW){      maskCol<-sample(nCOL, nMask)      randval<-runif(length(maskCol))      ixmin<-randval<0.5 ixmax<-randval>=0.5 ...