R Create Data.Frame Like Read.Csv

In exploring using R.NET and RserveLink inside of C#, I ran into a couple performance issues with my scenario. In load balanced environments which will need to pull the random forests and the cached CSV data saving the files to a network share makes sense, but there is a cost. My initial process was:

1. In C#, load and prep the data.
2. Serialize the data into CSV format and save to a network share.
3. Use R.NET or RServeLink, send the below commands to pull the Random Forest and the CSV data and run the data through the Random Forest.

Below is the proof of concept R code.

library("randomForest")
library("caret")

mydata = read.csv(file="IntPonAllTheData.csv",head=TRUE,row.names="IntPonID")

test.predict <- predict(readRDS('//intponsrv/Data/RandomForest/CLASS123.rf'), mydata)
write.table(test.predict)

I was able to combine several of the lines, but it didn't improve the performance very much. However, if I could remove 2 of the network and file I/O trips, that would greatly improve the performance. The only question was, how do I create a data.frame in R that would produce the same object as read.csv. I inquired to the #R freenode channel (they are awesome, check them out and stay a while) and they suggested that I look into the save function or dputs function. I was able to use dputs and serialize the data.frame. The format of the serialization wasn't an exact match, but it was close enough that I could figure out how the data.frame is structured relative to the CSV data. The following is my converted code which generates a data.frame directly.

library("randomForest")
library("caret")

df <- data.frame(Q1 = c(0.301775147928994,0.301775147928994,0.301775147928994,0.301775147928994),Q2 = c(0.301775147928994,0.301775147928994,0.301775147928994,0.301775147928994),Q2 = c(0.094674556213018,0.094674556213018,0.094674556213018,0.094674556213018),Q3 = c(0.301775147928994,0.301775147928994,0.301775147928994,0.301775147928994),Q4 = c(0.082840236686391,0.082840236686391,0.082840236686391,0.082840236686391),row.names = c("baseline","TEST1","TEST2","TEST3"))
write.table(df)

test.predict <- predict(readRDS('//intponsrv/Data/RandomForest/CLASS123.rf'), df)
write.table(test.predict)

Instead of generating the CSV, I can generate the data.frame statement with the same result. After integrating it with my C# code, this produced a 60-80% improvement in processing time per prediction, when repeatedly processing large data sets.

RserveLink Eval Failed Even When Successful

I have been playing with RserveLink and Rserve recently and found that it worked very well, except in one case (which is not really a fringe case) where it reliably returns an "Eval Failed" exception when the command actually succeeded. I was able to run the command in RStudio (Awesome R IDE) and RGUI successfully.

I am not going to include the R statement that generated the issue because it is large, but suffice it to say that it reliably works with much larger statements.

The response we get for the 4th index is 0 and the library jumps over the if statement. However, dragging the debug arrow onto the DataLength line and run the code, it continues without error.

if (response[4] != 0)
{
    Int32 DataLength = BitConverter.ToInt32(response, 4);

I changed the section of code converting the code to the following. Essentially, just getting the response data length and assume that a data length greater than 0 is successful.

Int32 DataLength = BitConverter.ToInt32(response, 4);
if (DataLength != 0)
{

The original code was posted in a zip file on SourceForge and posted back in 2007. I have emailed the author (awaiting reply) and posted my change on my RserveLink GitHub repository. The binary library targeting 4.0 Client and 4.5.1 can be found in the releases section.

I have also published an RserveLink Nuget package for ease of adding to .Net projects.