Create R Model

Updated: April 26, 2017

Creates an R model using custom resources

Category: Data Transformation / Manipulation

You can use the Create R Model module to create an untrained model from R script that you provide. You can base your model on any learner that is included in an R package in the Azure Machine Learning environment.

After you create the model, you can use Train Model to train the model on a dataset, like any other learner in Azure Machine Learning. The trained model can be passed to Score Model to use the model to make predictions. The trained model can then be saved, and the scoring workflow can be published as a web service.

Warning
Currently it is not possible to pass the scored results of an R model to Evaluate Model or Cross-Validate Model. If you need to evaluate a model, you can write custom R script and run it using the Execute R Script module.

In addition to using the Create R Model to save and re-use custom R modules, you can create your own implementation of a modeling and data management process using R, upload the files in zipped format to your workspace, and then register the package as a custom module. For more information, see Custom R Modules.

To illustrate the process, this example implements a two-class Naïve Bayes classifier by using the e1070 package. However, you can use any type of learner that is included in the many R packages supported in Azure Machine Learning.

Tip
See this sample in the Cortana Intelligence Gallery. You can also copy the example to your workspace and make changes! Create R Model

To get started, you must provide two user-defined scripts as inputs:

• Training script. The R script that you provide here is used to train the model. When you run the experiment, it is deployed to the Train Model module.

• Scoring script. The R script that you provide on this input is for scoring only. when you run the experiment, it is deployed to and runs using the Score Model module.

Training Script

The following example demonstrates the type of code you might use in the training script. It loads an R package, creates model using a learner from the package, and configures the feature and label columns using the predefined constants and functions provided in Create R Model.

library(e1071)   
features <- get.feature.columns(dataset)   
labels   <- as.factor(get.label.column(dataset))   
train.data <- data.frame(features, labels)   
feature.names <- get.feature.column.names(dataset)   
names(train.data) <- c(feature.names, "Class")   
model <- naiveBayes(Class ~ ., train.data)   

Step-by-step explanation of code

1. The first line loads the R package, e1071, which contain the Naïve Bayes classifier algorithm we want to use.

   Copy

   library(e1071)  
   
   

   Since this is one of the packages pre-installed in the Azure Machine Learning environment, you don’t need to download or install the package.

   For information about how to get the full list of supported packages, see R Language Modules.

2. The next three lines get the feature columns and the label column from the dataset, and combine them into a new R data frame that is named train.data:

   Copy

   features <- get.feature.columns(dataset)   
   labels   <- as.factor(get.label.column(dataset))   
   train.data <- data.frame(features, labels)   
   
   

   The predefined function, get.label.columns(), returns the column that is selected as the class label in the Train Model module.

3. The predefined function, get.feature.columns(), selects the columns that were designated as features in the metadata for dataset.

   Warning
   By default, all columns except the label column are considered features in Studio. If you want to select specific columns as features, use Edit Metadata, or select a set of columns within the R script.

4. In the fifth and sixth lines, the predefined function, get.feature.column.names(dataset), is used to get feature column names from the dataset. Those names are designated as the names for columns in train.data, and a temporary name Class is created for the label column.

   Copy

   feature.names <- get.feature.column.names(dataset)   
   names(train.data) <- c(feature.names, "Class")   
   
   

5. The final line of the code trains the Naïve Bayes classifier algorithm by using the labels and features in the train.data data frame.

   Copy

   model <- naiveBayes(Class ~ ., train.data)    
   
   

   Warning
   The model must always be assigned to a variable called model in both the train and score scripts.

Scoring Script

The following code illustrates the type of R code that you would provide for your scoring script.

library(e1071)   
probabilities <- predict(model, dataset, type="raw")[,2]   
classes <- as.factor(as.numeric(probabilities >= 0.5))   
scores <- data.frame(classes, probabilities)   

Step-by-step explanation of code

1. The first line loads the package.

   Copy

   library(e1071)  
   
   

2. The second line computes the predicted probabilities for the scoring dataset by using the trained model from the training script.

   Copy

   probabilities <- predict(model, dataset, type="raw")[,2]   
   
   

3. The third line applies a threshold of 0.5 to probabilities when assigning the predicted class labels.

   Copy

   classes <- as.factor(as.numeric(probabilities >= 0.5))   
   
   

4. The final line combines the class labels and probabilities into the output data frame, scores.

   Copy

   scores <- data.frame(classes, probabilities)   
   
   

   Warning
   The output dataframe must have the name scores.

Publishing the R Model as a Web Service

After you have run the experiment by using the R model, you can publish the experiment as a web service.

1. Click Create Scoring Experiment in the experiment editor.

   The trained R model is saved, and the training experiment graph is transformed into a scoring experiment.

2. Run the scoring experiment at least once.

3. Click Publish Web Service.

Note
By default, the web service expects all input columns from the training data to be provided, including the label column. You can add an instance of Select Columns in Dataset between the input data source and the Score Model module to exclude the label you’re trying to predict.

• The Create R Model module supports use of CRAN R only. You cannot select another version of R, or use Microsoft R Open.

• The model is cached after the first run of the module and the module is not invoked in subsequent runs until any changes in input scripts are done. Please take this behavior into consideration if your R scripts use any of the following:

  • Functions that generate random numbers

  • Functions that generate random numbers

  • Other nondeterministic functions

• R models cannot be used with these modules: Tune Model Hyperparameters, Cross-Validate Model, One-vs-All Multiclass or Ordinal Regression.

• R models do not automatically perform feature normalization of categorical data or handle missing values. Handling of such variables should be done within the training and scoring R scripts.

For examples of how to use this module in machine learning experiments, see the Cortana Intelligence Gallery:

• This experiment demonstrates how to use the Create R Model module to train, and score a naive bayes classification model using the breast cancer dataset. It uses the Execute Python Script to calculate performance and plot the performance curve: Create R Model

Execute R Script
Module Categories and Descriptions
R Language Modules