regtools (version 1.7.0)

krsFit: Tools for Neural Networks

Description

Tools to complement existing neural networks software, notably a more "R-like" wrapper to fitting data with R's keras package.

Usage

krsFit(x,y,hidden,acts=rep("relu",length(hidden)),learnRate=0.001,
   conv=NULL,xShape=NULL,classif=TRUE,nClass=NULL,nEpoch=30,
   scaleX=TRUE,scaleY=TRUE)
krsFitImg(x,y,hidden=c(100,100),acts=rep("relu",length(hidden)),
    nClass,nEpoch=30) 
# S3 method for krsFit
predict(object,...)
diagNeural(krsFitOut)

Arguments

object

An object of class 'krsFit'.

...

Data points to be predicted, 'newx'.

x

X data, predictors, one row per data point, in the training set. Must be a matrix.

y

Numeric vector of Y values. In classification case must be integers, not an R factor, and take on the values 0,1,2,..., nClass-1

hidden

Vector of number of units per hidden layer, or the rate for a dropout layer.

acts

Vector of names of the activation functions, one per hidden layer. Choices inclde 'relu', 'sigmoid', 'tanh', 'softmax', 'elu', 'selu'.

learnRate

Learning rate.

conv

R list specifying the convolutional layers, if any.

xShape

Vector giving the number of rows and columns, and in the convolutional case with multiple channels, the number of channels.

classif

If TRUE, indicates a classification problem.

nClass

Number of classes.

nEpoch

Number of epochs.

krsFitOut

An object returned by krstFit.

scaleX

If TRUE, scale X columns.

scaleY

If TRUE, scale Y columns.

Details

The krstFit function is a wrapper for the entire pipeline in fitting the R keras package to a dataset: Defining the model, compiling, stating the inputs and so on. As a result, the wrapper allows the user to skip those details (or not need to even know them), and define the model in a manner more familiar to R users.

The paired predict.krsFit takes as its first argument the output of krstFit, and newx, the points to be predicted.

Examples

# NOT RUN {
# }
# NOT RUN {
library(keras)
data(peDumms) 
ped <- peDumms[,c(1,20,22:27,29,32,31)]
# predict wage income
x <- ped[,-11] 
y <- ped[,11] 
z <- krsFit(x,y,c(50,50,50),classif=FALSE,nEpoch=25) 
preds <- predict(z,x) 
mean(abs(preds-y))  # something like 25000

x <- ped[,-(4:8)] 
y <- ped[,4:8] 
y <- dummiesToInt(y,FALSE) - 1
z <- krsFit(x,y,c(50,50,0.20,50),classif=TRUE,nEpoch=175,nClass=6) 
preds <- predict(z,x)
mean(preds == y)   # something like 0.39

# obtain MNIST training and test sets; the following then uses the
# example network of 

# https://databricks-prod-cloudfront.cloud.databricks.com/
# public/4027ec902e239c93eaaa8714f173bcfc/2961012104553482/
# 4462572393058129/1806228006848429/latest.html

# converted to use the krsFit wrapper

x <- mntrn[,-785] / 255 
y <- mntrn[,785]
xShape <- c(28,28)

# define convolutional layers
conv1 <- list(type='conv2d',filters=32,kern=3)
conv2 <- list(type='pool',kern=2)
conv3 <- list(type='conv2d',filters=64,kern=3) 
conv4 <- list(type='pool',kern=2)
conv5 <- list(type='drop',drop=0.5)

# call wrapper, 1 dense hidden layer of 128 units, then dropout layer
# with proportion 0.5
z <- krsFit(x,y,conv=list(conv1,conv2,conv3,conv4,conv5),c(128,0.5),
   classif=TRUE,nClass=10,nEpoch=10,xShape=c(28,28),scaleX=FALSE,scaleY=FALSE)

# try on test set
preds <- predict(z,mntst[,-785]/255)
mean(preds == mntst[,785])  # 0.98 in my sample run

# }
# NOT RUN {
# }