R/feedForward.R

In deepdive: Deep Learning for General Purpose

#' FeedForward for deepnet

#' @return
#' @noRd

feedForward <- function(xBiased, weightMatrix, activation,reluLeak, modelType,baisUnits) {



  a_output = list()
  z_in = list()


  for (i in 1:length(weightMatrix)) {

    if (i == 1) {

      a_input<-xBiased

      z_input <- as.matrix(a_input) %*% weightMatrix[[i]]
    }else{



      z_input <- as.matrix(a_input) %*% weightMatrix[[i]]+
        matrix(rep(t(baisUnits[[i]]),nrow(a_input)),nrow=nrow(a_input),byrow = T)


    }





      if (i < length(weightMatrix)) {

        if(activation[i]=="relu"){


          a_input <- ifelse(z_input < 0, reluLeak, z_input)
          }else if( activation[i]=="none"){

            a_input<-z_input

          }else if (activation[i]=="sigmoid"){
            a_input<-1/(1+exp(-z_input))


          }else if (activation[i]=="sin"){
            a_input<-sin(z_input)
          }else if (activation[i]=="cos"){
            a_input<-cos(z_input)


          }

      } else{


        if(modelType=="regress"){
        a_input <- z_input}else if(modelType=="binary"){

        a_input<-(1/(1+exp(-z_input)))

        }else if(modelType=="multiClass"){


          a_input<-  exp(z_input)/rowSums(exp(z_input))
        }

      }





    a_output[[i]] <- a_input
    z_in[[i]] <- z_input
  }




  return(list(a_output = a_output,
              z_in = z_in))
}