In mertayD/CodingProject3: coding project Neural Nets
knitr::opts_chunk$set(
collapse = TRUE,
comment = "#>"
)
usethis::use_package("CodingProject3")
usethis::use_package("ggplot2")
data.name.vec <- c(
"spam",
"zip.train",
"prostate",
"ozone")
data.list <- list()
library(ElemStatLearn)
for(data.name in data.name.vec)
{
data(list = data.name, package="ElemStatLearn")
data.list[[data.name]] <- get(data.name)
}
is.binary <- ElemStatLearn::zip.train[,1] %in% c(0,1)
data.list <- list(
spam=list(
label.vec=ifelse(ElemStatLearn::spam$spam=="spam", 1, 0),
feature.mat=as.matrix(ElemStatLearn::spam[, 1:57])),
SAheart=list(
label.vec=ifelse(ElemStatLearn::SAheart$SAheart=="chd", 1, 0),
feature.mat=as.matrix(ElemStatLearn::SAheart[, -ncol(ElemStatLearn::SAheart)])),
zip.train=list(
label.vec=ElemStatLearn::zip.train[is.binary,1],
feature.mat=ElemStatLearn::zip.train[is.binary,-1]),
prostate=list(
label.vec=as.numeric(ElemStatLearn::prostate$lpsa),
feature.mat=as.matrix(ElemStatLearn::prostate[,-c(9, 10)])),
ozone=list(
label.vec=ElemStatLearn::ozone$ozone,
feature.mat=as.matrix(ElemStatLearn::ozone[,-1]))
)
n.folds <- 4
resultsList <- list()
baseline <- list()
mean.loss.list <- list()
#current.set.matrix <-
str(data.list)
for(data.name in names(data.list))
{
one.data.set <- data.list[[data.name]]
set.seed(1)
# print("This is one.data.set")
# print(one.data.set)
# print(data.list[[data.name]])
# print(names(data.list))
#print(one.data.set$feature.mat[1:nrow(one.data.set$feature.mat)])
#print(one.data.set[1:nrow(one.data.set$feature.mat), -1])
# variables for function
if(data.name == 'ozone'){
X.mat <- one.data.set$feature.mat
y.vec <- one.data.set$label.vec
fold.vec <- sample(rep(1:n.folds, l=nrow(X.mat)))
max.ite <- 50
result.mat.list <- list()
for(test.fold in 1:length(unique(fold.vec)))
{
is.test <- fold.vec == test.fold
is.train <- !is.test
fold_data <- which(fold.vec == test.fold)
X.train <- X.mat[-fold_data ,]
X.valid <- X.mat[fold_data ,]
y.train <- y.vec[-fold_data]
y.valid <- y.vec[fold_data]
fold.vec.train <- sample(rep(1:n.folds, l=nrow(X.train)))
#For each train/test split, to show that your algorithm is actually learning something non-trivial from the inputs/features, compute a baseline predictor that ignores the inputs/features.
#Regression: the mean of the training labels/outputs.
baseline <- mean(y.train)
# print(data.name)
# print(data.list[[data.name]])
fold.v <- sample(rep(1:4, l=nrow(X.train)))
resultES <- CodingProject3::NNetEarlyStoppingCV(X.train,
y.train,
fold.v,
max.iterations=max.ite,
step.size=0.05,
n.hidden.units=30)
}
# pred.prob.list[test.fold] <- list(
# earlyStopping=resultES$predict(one.data.set$feature.mat[is.test,]),
# #L2regularized=resultREG$predict(one.data.set$feature.mat[is.test,]),
# baseline=rep(baseline, sum(is.test)))
# For each data set, compute a 4 x 3 matrix of mean test loss values:
# each of the four rows are for a specific test set,
# the first column is for the early stopping predictor,
# the second column is for the L2 regularized predictor,
# the third column is for the baseline/un-informed predictor.
# mat <- cbind(resultES$penalty.vec, result$penalty.vec)
# str(pred.prob.list)
# test.fold.result.list <- list()
# for(algo in names(pred.prob.list))
# {
# pred.prob.vec <- pred.prob.list[[algo]]
# pred.class.vec <- ifelse(pred.prob.vec > 0.5, 1, 0)
# test.fold.result.list[[algo]] <- mean(data.set$labels[is.test] != #pred.class.vec)
# }
}
}
# For spam SAheart and train
# else
# {
#For each train/test split, to show that your algorithm is actually learning something non-trivial from the inputs/features, compute a baseline predictor that ignores the inputs/features.
# Binary classification: the most frequent class/label/output in the training data.
#prdictionBase <- mfv(fold.vec)
#baseline.append(mfv(fold.vec))
# resultES <- LMLogisticLossEarlyStoppingCV(X.mat,
# y.vec,
# fold.vec,
#max.iterations)
# result <- LMLogisticLossIterations(X.mat,
# y.vec,
# fold.vec,
# max.iterations)
# resultList.append(resultES)
# resultList.append(result)
# }
# For each data set, compute a 4 x 3 matrix of mean test loss values:
# each of the four rows are for a specific test set,
# the first column is for the early stopping predictor,
# the second column is for the L2 regularized predictor,
# the third column is for the baseline/un-informed predictor.
#mat <- cbind(resultES$penalty.vec, result$penalty.vec)
# List is (LMLogisticLossEarlyStoppingCV, LMLogisticLossIterations for spam SAheart and then train,
# Then LMSquareLossEarlyStoppingCV LMSquareLossIterations for prostate and ozone)
# ggplot(resultsList[5]$train.loss.vec)
# ggplot(resultsList[5]$validation.loss.vec)
# ggplot(resultsList[6]$train.loss.vec)
# ggplot(resultsList[6]$validation.loss.vec)
Data set 1: spam
Matrix of loss values
#print out and/or plot the matrix of loss values
# pander::pandoc.table(resultsList[1]$train.loss.mat)
# pander::pandoc.table(resultsList[2]$train.loss.mat)
comment on difference between NN and baseline.
Train/validation loss plot
#plot the two loss functions.
plot(resultES$mean.train.loss.vec)
plot(resultES$mean.validation.loss)
# ggplot(resultsList[2]$train.loss.vec)
# ggplot(resultsList[2]$validation.loss.vec)
What are the optimal regularization parameters?
# selected
Data set 2: SAheart
Matrix of loss values
#print out and/or plot the matrix.
# pander::pandoc.table(resultsList[3]$train.loss.mat)
# pander::pandoc.table(resultsList[4]$train.loss.mat)
comment on difference between NN and baseline.
Train/validation loss plot
#plot the two loss functions.
# ggplot(resultsList[3]$train.loss.vec)
# ggplot(resultsList[3]$validation.loss.vec)
# ggplot(resultsList[4]$train.loss.vec)
# ggplot(resultsList[4]$validation.loss.vec)
What are the optimal regularization parameters?
# selected
Data set 3: zip.train
Matrix of loss values
#print out and/or plot the matrix.
# pander::pandoc.table(resultsList[5]$train.loss.mat)
# pander::pandoc.table(resultsList[6]$train.loss.mat)
comment on difference between NN and baseline.
Train/validation loss plot
#plot the two loss functions.
#plot the two loss functions.
# ggplot(resultsList[5]$train.loss.vec)
# ggplot(resultsList[5]$validation.loss.vec)
# ggplot(resultsList[6]$train.loss.vec)
# ggplot(resultsList[6]$validation.loss.vec)
What are the optimal regularization parameters?
# selected
Data set 4: prostate
Matrix of loss values
#print out and/or plot the matrix.
# pander::pandoc.table(resultsList[7]$train.loss.mat)
# pander::pandoc.table(resultsList[8]$train.loss.mat)
comment on difference between NN and baseline.
Train/validation loss plot
#plot the two loss functions.
# ggplot(resultsList[7]$train.loss.vec)
# ggplot(resultsList[7]$validation.loss.vec)
# ggplot(resultsList[8]$train.loss.vec)
# ggplot(resultsList[8]$validation.loss.vec)
What are the optimal regularization parameters?
# selected
Data set 5: ozone
Matrix of loss values
#print out and/or plot the matrix.
# pander::pandoc.table(resultsList[9]$train.loss.mat)
# pander::pandoc.table(resultsList[10]$train.loss.mat)
comment on difference between NN and baseline.
Train/validation loss plot
#plot the two loss functions.
# ggplot(resultsList[9]$train.loss.vec)
# ggplot(resultsList[9]$validation.loss.vec)
# ggplot(resultsList[10]$train.loss.vec)
# ggplot(resultsList[10]$validation.loss.vec)
What are the optimal regularization parameters?
Loss
$L(w,v) = \frac{1}2 \times [\Sigma^{n}{1}[w^{t}S[V^{T}x{i}] - y_{i}]^2]$
Gradient/Backprop
- $A = XV$
- $Z = S[A]$
- $b = ZW$
- $\delta^{w} = b - y$
- $\delta^{v} = Diag(\delta^{v}) \times S'[A] \times Diag(w)$
- $\nabla_{w}L(w,v) = (Z^{t} \delta^{w}) / n$
- $\nabla_{v}L(w,v) = (X^{t} \delta^{v}) / n$
mertayD/CodingProject3 documentation built on May 14, 2019, 3:07 a.m.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
knitr::opts_chunk$set( collapse = TRUE, comment = "#>" )
usethis::use_package("CodingProject3") usethis::use_package("ggplot2") data.name.vec <- c( "spam", "zip.train", "prostate", "ozone") data.list <- list() library(ElemStatLearn) for(data.name in data.name.vec) { data(list = data.name, package="ElemStatLearn") data.list[[data.name]] <- get(data.name) } is.binary <- ElemStatLearn::zip.train[,1] %in% c(0,1) data.list <- list( spam=list( label.vec=ifelse(ElemStatLearn::spam$spam=="spam", 1, 0), feature.mat=as.matrix(ElemStatLearn::spam[, 1:57])), SAheart=list( label.vec=ifelse(ElemStatLearn::SAheart$SAheart=="chd", 1, 0), feature.mat=as.matrix(ElemStatLearn::SAheart[, -ncol(ElemStatLearn::SAheart)])), zip.train=list( label.vec=ElemStatLearn::zip.train[is.binary,1], feature.mat=ElemStatLearn::zip.train[is.binary,-1]), prostate=list( label.vec=as.numeric(ElemStatLearn::prostate$lpsa), feature.mat=as.matrix(ElemStatLearn::prostate[,-c(9, 10)])), ozone=list( label.vec=ElemStatLearn::ozone$ozone, feature.mat=as.matrix(ElemStatLearn::ozone[,-1])) ) n.folds <- 4 resultsList <- list() baseline <- list() mean.loss.list <- list() #current.set.matrix <- str(data.list) for(data.name in names(data.list)) { one.data.set <- data.list[[data.name]] set.seed(1) # print("This is one.data.set") # print(one.data.set) # print(data.list[[data.name]]) # print(names(data.list)) #print(one.data.set$feature.mat[1:nrow(one.data.set$feature.mat)]) #print(one.data.set[1:nrow(one.data.set$feature.mat), -1]) # variables for function if(data.name == 'ozone'){ X.mat <- one.data.set$feature.mat y.vec <- one.data.set$label.vec fold.vec <- sample(rep(1:n.folds, l=nrow(X.mat))) max.ite <- 50 result.mat.list <- list() for(test.fold in 1:length(unique(fold.vec))) { is.test <- fold.vec == test.fold is.train <- !is.test fold_data <- which(fold.vec == test.fold) X.train <- X.mat[-fold_data ,] X.valid <- X.mat[fold_data ,] y.train <- y.vec[-fold_data] y.valid <- y.vec[fold_data] fold.vec.train <- sample(rep(1:n.folds, l=nrow(X.train))) #For each train/test split, to show that your algorithm is actually learning something non-trivial from the inputs/features, compute a baseline predictor that ignores the inputs/features. #Regression: the mean of the training labels/outputs. baseline <- mean(y.train) # print(data.name) # print(data.list[[data.name]]) fold.v <- sample(rep(1:4, l=nrow(X.train))) resultES <- CodingProject3::NNetEarlyStoppingCV(X.train, y.train, fold.v, max.iterations=max.ite, step.size=0.05, n.hidden.units=30) } # pred.prob.list[test.fold] <- list( # earlyStopping=resultES$predict(one.data.set$feature.mat[is.test,]), # #L2regularized=resultREG$predict(one.data.set$feature.mat[is.test,]), # baseline=rep(baseline, sum(is.test))) # For each data set, compute a 4 x 3 matrix of mean test loss values: # each of the four rows are for a specific test set, # the first column is for the early stopping predictor, # the second column is for the L2 regularized predictor, # the third column is for the baseline/un-informed predictor. # mat <- cbind(resultES$penalty.vec, result$penalty.vec) # str(pred.prob.list) # test.fold.result.list <- list() # for(algo in names(pred.prob.list)) # { # pred.prob.vec <- pred.prob.list[[algo]] # pred.class.vec <- ifelse(pred.prob.vec > 0.5, 1, 0) # test.fold.result.list[[algo]] <- mean(data.set$labels[is.test] != #pred.class.vec) # } } } # For spam SAheart and train # else # { #For each train/test split, to show that your algorithm is actually learning something non-trivial from the inputs/features, compute a baseline predictor that ignores the inputs/features. # Binary classification: the most frequent class/label/output in the training data. #prdictionBase <- mfv(fold.vec) #baseline.append(mfv(fold.vec)) # resultES <- LMLogisticLossEarlyStoppingCV(X.mat, # y.vec, # fold.vec, #max.iterations) # result <- LMLogisticLossIterations(X.mat, # y.vec, # fold.vec, # max.iterations) # resultList.append(resultES) # resultList.append(result) # } # For each data set, compute a 4 x 3 matrix of mean test loss values: # each of the four rows are for a specific test set, # the first column is for the early stopping predictor, # the second column is for the L2 regularized predictor, # the third column is for the baseline/un-informed predictor. #mat <- cbind(resultES$penalty.vec, result$penalty.vec) # List is (LMLogisticLossEarlyStoppingCV, LMLogisticLossIterations for spam SAheart and then train, # Then LMSquareLossEarlyStoppingCV LMSquareLossIterations for prostate and ozone)
# ggplot(resultsList[5]$train.loss.vec) # ggplot(resultsList[5]$validation.loss.vec) # ggplot(resultsList[6]$train.loss.vec) # ggplot(resultsList[6]$validation.loss.vec)
Data set 1: spam
Matrix of loss values
#print out and/or plot the matrix of loss values # pander::pandoc.table(resultsList[1]$train.loss.mat) # pander::pandoc.table(resultsList[2]$train.loss.mat)
comment on difference between NN and baseline.
Train/validation loss plot
#plot the two loss functions. plot(resultES$mean.train.loss.vec) plot(resultES$mean.validation.loss) # ggplot(resultsList[2]$train.loss.vec) # ggplot(resultsList[2]$validation.loss.vec)
What are the optimal regularization parameters?
# selected
Data set 2: SAheart
Matrix of loss values
#print out and/or plot the matrix. # pander::pandoc.table(resultsList[3]$train.loss.mat) # pander::pandoc.table(resultsList[4]$train.loss.mat)
comment on difference between NN and baseline.
Train/validation loss plot
#plot the two loss functions. # ggplot(resultsList[3]$train.loss.vec) # ggplot(resultsList[3]$validation.loss.vec) # ggplot(resultsList[4]$train.loss.vec) # ggplot(resultsList[4]$validation.loss.vec)
What are the optimal regularization parameters?
# selected
Data set 3: zip.train
Matrix of loss values
#print out and/or plot the matrix. # pander::pandoc.table(resultsList[5]$train.loss.mat) # pander::pandoc.table(resultsList[6]$train.loss.mat)
comment on difference between NN and baseline.
Train/validation loss plot
#plot the two loss functions. #plot the two loss functions. # ggplot(resultsList[5]$train.loss.vec) # ggplot(resultsList[5]$validation.loss.vec) # ggplot(resultsList[6]$train.loss.vec) # ggplot(resultsList[6]$validation.loss.vec)
What are the optimal regularization parameters?
# selected
Data set 4: prostate
Matrix of loss values
#print out and/or plot the matrix. # pander::pandoc.table(resultsList[7]$train.loss.mat) # pander::pandoc.table(resultsList[8]$train.loss.mat)
comment on difference between NN and baseline.
Train/validation loss plot
#plot the two loss functions. # ggplot(resultsList[7]$train.loss.vec) # ggplot(resultsList[7]$validation.loss.vec) # ggplot(resultsList[8]$train.loss.vec) # ggplot(resultsList[8]$validation.loss.vec)
What are the optimal regularization parameters?
# selected
Data set 5: ozone
Matrix of loss values
#print out and/or plot the matrix. # pander::pandoc.table(resultsList[9]$train.loss.mat) # pander::pandoc.table(resultsList[10]$train.loss.mat)
comment on difference between NN and baseline.
Train/validation loss plot
#plot the two loss functions. # ggplot(resultsList[9]$train.loss.vec) # ggplot(resultsList[9]$validation.loss.vec) # ggplot(resultsList[10]$train.loss.vec) # ggplot(resultsList[10]$validation.loss.vec)
What are the optimal regularization parameters?
Loss
$L(w,v) = \frac{1}2 \times [\Sigma^{n}{1}[w^{t}S[V^{T}x{i}] - y_{i}]^2]$
Gradient/Backprop
- $A = XV$
- $Z = S[A]$
- $b = ZW$
- $\delta^{w} = b - y$
- $\delta^{v} = Diag(\delta^{v}) \times S'[A] \times Diag(w)$
- $\nabla_{w}L(w,v) = (Z^{t} \delta^{w}) / n$
- $\nabla_{v}L(w,v) = (X^{t} \delta^{v}) / n$
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Embedding an R snippet on your website
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.