10.3 Lab: A Single Layer Network on the Hitters Data

How to perform deep learning in RStudio:

Single Layer

library(ISLR2)
Gitters <- na.omit(Hitters)
n <- nrow(Gitters)

set.seed(13)
ntest <- trunc(n / 3) #rounding numbers
testid <- sample(1:n, ntest)

Fit the model with testid selection

training <- Gitters[-testid, ]
testing <- Gitters[testid, ]

lfit <- lm(Salary ~ ., data = training)

lpred <- predict(lfit, testing)

pred_test <- cbind(testing,lpred)

mean(abs(pred_test$lpred-pred_test$Salary))

## [1] 254.6687

Standardize the matrix and fit the lasso using glmnet

x <- scale(model.matrix(Salary ~ . - 1, data = Gitters))

y <- Gitters$Salary

library(glmnet)

## Loading required package: Matrix

## 
## Attaching package: 'Matrix'

## The following objects are masked from 'package:tidyr':
## 
##     expand, pack, unpack

## Loaded glmnet 4.1-8

cvfit <- cv.glmnet(x[-testid, ], y[-testid], type.measure = "mae")
cpred <- predict(cvfit, x[testid, ], s = "lambda.min")
mean(abs(y[testid] - cpred))

## [1] 252.2994

There are two ways to fit the Neural Network:

using {keras} implies pyton
using {torch}

book-Lab-source

Fit the neural network with {keras}:

keras_model_sequential()
layer_dense()
layer-dropout()

Keras requires some installation on RStudio:

library(ISLR2)
tryCatch(
  remove.packages(c("keras", "tensorflow", "reticulate")),
  error = function(e) "Some or all packages not previously installed, that's ok!"
)

install.packages("keras", repos = 'https://cloud.r-project.org')

write('RETICULATE_AUTOCONFIGURE=FALSE', file = "~/.Renviron", append = TRUE)
write(sprintf('RETICULATE_MINICONDA_PATH=%s',
           normalizePath("~/islr-miniconda", winslash = "/", mustWork = FALSE)),
   file = "~/.Renviron", append = TRUE)

# restart R

source(system.file("helpers", "install.R", package = "ISLR2"))

reticulate::install_miniconda(force = TRUE)
tensorflow::install_tensorflow()

library(tidyverse)
library(keras)
library(tensorflow)

modnn <- keras_model_sequential() %>%
     layer_dense(units = 50, activation = "relu",
        input_shape = ncol(x)) %>%
     layer_dropout(rate = 0.4) %>%
     layer_dense(units = 1)

Resources:

example-tensorflow

The second way is to use {torch}:

library(torch)
library(luz) # high-level interface for torch
library(torchvision) # for datasets and image transformation
library(torchdatasets) # for datasets we are going to use
library(zeallot)
torch_manual_seed(13)

modnn <- nn_module(
  initialize = function(input_size) {
    self$hidden <- nn_linear(input_size, 50)
    self$activation <- nn_relu()
    self$dropout <- nn_dropout(0.4)
    self$output <- nn_linear(50, 1)
  },
  forward = function(x) {
    x %>% 
      self$hidden() %>% 
      self$activation() %>% 
      self$dropout() %>% 
      self$output()
  }
)

Torch Lab - html