**This is an old erroneous test which I attempted without much knowledge of L1, L2 regression and benchmarking. I attempted this test again after a lot of research and my new test is in the same repo with the name easy_test.md.
Now it's only a memoir of the progress that I underwent throughout the week while attempting the tests
Including the required library files and initializing empty variables
> rm(list = ls())
> library(reshape2)
> library(glmnet)
> library(iregnet)
> library(ggplot2)
> library(microbenchmark)
> #Prostate dataset
> data("Prostate", package = "lasso2")
> #Benchmark times for glmnet and irgnet respectively
> glm_time <- vector()
> irg_time <- vector()
> #The data set size vector
> ds_size <- vector()
Checking the difference between the results produced by irgnet and glmnet for a dataset size of 50
> #Reshaping as matrix
> X = as.matrix(Prostate[1:50, c(2:9)])
> Y = matrix(c(Prostate[1:50, 1]))
> dimnames(Y) <- NULL
> dimnames(X) <- NULL
> #Setting the minimum fractional change in deviance for stoppinng path to 0 so that function
> #continues all along the path, even without much change(effectively, 100 lambda values are obtained)
> glmnet.control(fdev = 0)
> #Building the model with glmnet(Lasso is selected by default)
> lasso_glm <- glmnet(y = Y, x = X, family="gaussian")
> #Since, the input to iregnet cannot be a one dimensional matrix, replicating column one into two
> Y = matrix(c(Prostate[1:50, 1],Prostate[1:50, 1]), nrow = 50, ncol = 2)
> dimnames(Y) <- NULL
> #Building the model with irgnet(Lasso selected by default)
> lasso_irg <- iregnet(y = Y, x = X, family="gaussian")
> #Plotting
> plot(lasso_irg)
> plot(lasso_glm)
> #Finding the change in lambda values obtained
> change <- lasso_irg$lambda - lasso_glm$lambda
> summary(change)
Min. 1st Qu. Median Mean 3rd Qu. Max.
-3.332e-02 -3.335e-03 -3.336e-04 -3.751e-03 -3.660e-05 -3.660e-06
We see that the mean change of lambda over the two functions is -3.751 x 10^(-3), i.e. can be considered "almost similar" upto a certain degree of error.
Even the lasso regression plots of the two functions are almost identical.
Plot obtained for lasso_irg
Plot obtained for lasso_glm
Now,
Evaluating the runtime of both the functions using the package microbenchmark over the data set size from 50 to 97
> #Iterating i from 1 to 47 and then adding 50 to it
> for(i in 1:47)
+ {
+ #Creating the Y and X matrices
+ Y = as.matrix(Prostate[1:50+i, 1])
+ X = as.matrix(Prostate[1:50+i, c(2:9)])
+ dimnames(Y) <- NULL
+ dimnames(X) <- NULL
+ #Setting the minimum fractional change in deviance for stoppinng path to 0
+ glmnet.control(fdev = 0)
+ #Timing the glmnet function using microbenchmark
+ mlm <- microbenchmark(
+ glmnet(y = Y, x = X, family="gaussian")
+ )
+ #Microbenchmark executes 100 times to evaluate runtime and hence taking mean of all the observation
+ # and dividing by 1000000 to obtain the time in milliseconds and storing in a vector
+ glm_time <- c(glm_time, mean(mlm$time)/1000000)
+ #Replicating column one into two for irgnet input
+ Y = matrix(c(Prostate[1:50+i, 1],Prostate[1:50+i, 1]), nrow = 50, ncol = 2)
+ dimnames(Y) <- NULL
+ #Timing irgnet using microbenchmark
+ mlm <- microbenchmark(
+ iregnet(y = Y, x = X, family="gaussian")
+ )
+ irg_time <- c(irg_time, mean(mlm$time)/1000000)
+ }
> #Defining the dataset sizse for which the benchmarks were performed
> ds_size <- c(51:97)
> #Storing the observations in a data frame
> df <- data.frame(ds_size, glm_time, irg_time)
> names(df) <- c("ds_size", "glmnet", "irgnet")
> #Melting the dataframe to make it ggplot friendly
> df_long <- melt(df, id = "ds_size")
> #Plotting the two runtimes using ggplot
> p <- ggplot(df_long, aes(x = ds_size, y = value, color = variable)) +
+ geom_line() +
+ ggtitle('Runtime(in milliseconds) vs Dataset Size') +
+ xlab('Dataset Size') +
+ ylab('Runtime (in milliseconds)')
> direct.label(p,"angled.boxes")
The resulting plot obtained
#Including the required libraries
rm(list = ls())
library(reshape2)
library(glmnet)
library(directlabels)
library(iregnet)
library(ggplot2)
library(microbenchmark)
#Prostate dataset
data("Prostate", package = "lasso2")
#Benchmark times for glmnet and irgnet respectively
glm_time <- vector()
irg_time <- vector()
#The data set size vector
ds_size <- vector()
#-----------------------------------------------------------------
#Reshaping as matrix
X = as.matrix(Prostate[1:50, c(2:9)])
Y = matrix(c(Prostate[1:50, 1]))
dimnames(Y) <- NULL
dimnames(X) <- NULL
#Setting the minimum fractional change in deviance for stoppinng path to 0 so that function
#continues all along the path, even without much change(effectively, 100 lambda values are obtained)
glmnet.control(fdev = 0)
#Building the model with glmnet(Lasso is selected by default)
lasso_glm <- glmnet(y = Y, x = X, family="gaussian")
#Since, the input to iregnet cannot be a one dimensional matrix, replicating column one into two
Y = matrix(c(Prostate[1:50, 1],Prostate[1:50, 1]), nrow = 50, ncol = 2)
dimnames(Y) <- NULL
#Building the model with irgnet(Lasso selected by default)
lasso_irg <- iregnet(y = Y, x = X, family="gaussian")
#Plotting
plot(lasso_irg)
plot(lasso_glm)
#Finding the change in lambda values obtained
change <- lasso_irg$lambda - lasso_glm$lambda
summary(change)
#-----------------------------------------------------------------
#Iterating i from 1 to 47 and then adding 50 to it
for(i in 1:47)
{
#Creating the Y and X matrices
Y = as.matrix(Prostate[1:50+i, 1])
X = as.matrix(Prostate[1:50+i, c(2:9)])
dimnames(Y) <- NULL
dimnames(X) <- NULL
#Setting the minimum fractional change in deviance for stoppinng path to 0
glmnet.control(fdev = 0)
#Timing the glmnet function using microbenchmark
mlm <- microbenchmark(
glmnet(y = Y, x = X, family="gaussian")
)
#Microbenchmark executes 100 times to evaluate runtime and hence taking mean of all the observation
# and dividing by 1000000 to obtain the time in milliseconds and storing in a vector
glm_time <- c(glm_time, mean(mlm$time)/1000000)
#Replicating column one into two for irgnet input
Y = matrix(c(Prostate[1:50+i, 1],Prostate[1:50+i, 1]), nrow = 50, ncol = 2)
dimnames(Y) <- NULL
#Timing irgnet using microbenchmark
mlm <- microbenchmark(
iregnet(y = Y, x = X, family="gaussian")
)
irg_time <- c(irg_time, mean(mlm$time)/1000000)
}
#Defining the dataset sizse for which the benchmarks were performed
ds_size <- c(51:97)
#Storing the observations in a data frame
df <- data.frame(ds_size, glm_time, irg_time)
names(df) <- c("ds_size", "glmnet", "irgnet")
#Melting the dataframe to make it ggplot friendly
df_long <- melt(df, id = "ds_size")
#Plotting the two runtimes using ggplot
p <- ggplot(df_long, aes(x = ds_size, y = value, color = variable)) +
geom_line() +
ggtitle('Runtime(in milliseconds) vs Dataset Size') +
xlab('Dataset Size') +
ylab('Runtime (in milliseconds)')
direct.label(p,"angled.boxes")