Converting R to PMML
3 likes2,997 views
The document discusses challenges in converting R models to PMML format and solutions using the r2pmml package. Specifically, it addresses issues like tightly coupled R scripts and model objects, lack of standardization, and inability to deploy models outside of R. The r2pmml package allows decorating model objects with supporting information and converting them to PMML, addressing these issues. This standardized format represents the R workflow and data using PMML structures, enabling "train once, deploy anywhere". The document provides examples of converting various R models like GLM and transforming features to PMML.
![R solution
Matrix interface ( ):
label = audit$Adjusted
features = audit[, !(colnames(audit) %in% c("Adjusted"))]
# Returns a `model` object
audit.model = model(x = features, y = label)
Formula interface ( ):
# Returns a `model.formula` object
audit.model = model(Adjusted ~ . + I(Income / (Hours * 52)), data = audit)](https://image.slidesharecdn.com/convertingrtopmml-171116200739/85/Converting-R-to-PMML-7-320.jpg)
![Bucketization (2/2)
<DerivedField name="cut(Income)" optype="categorical" dataType="string">
<Extension>cut(Income, breaks = 3)</Extension>
<Discretize field="Income">
<DiscretizeBin binValue="(129,1.61e+05]">
<Interval closure="openClosed" leftMargin="129.0" rightMargin="161000.0"/>
</DiscretizeBin>
<DiscretizeBin binValue="(1.61e+05,3.21e+05]">
<Interval closure="openClosed" leftMargin="161000.0" rightMargin="321000.0"/>
</DiscretizeBin>
<DiscretizeBin binValue="(3.21e+05,4.82e+05]">
<Interval closure="openClosed" leftMargin="321000.0" rightMargin="482000.0"/>
</DiscretizeBin>
</Discretize>
</DerivedField>](https://image.slidesharecdn.com/convertingrtopmml-171116200739/85/Converting-R-to-PMML-19-320.jpg)
Converting R to PMML
- 1. Converting R to PMML Villu Ruusmann Openscoring OÜ
- 2. "Train once, deploy anywhere"
- 3. R challenge "Fat code around thin model objects": ● Many packages to solve the same problem, many ways of doing and/or representing the same thing ● The unit of reusability is R script ● Model object is tightly coupled to the R script that produced it. Incomplete/split schemas ● No (widely adopted-) pipeline formalization
- 4. (J)PMML solution Evaluator evaluator = ...; List<InputField> argumentFields = evaluator.getInputFields(); List<ResultField> resultFields = Lists.union(evaluator.getTargetFields(), evaluator.getOutputFields()); Map<FieldName, ?> arguments = readRecord(argumentFields); Map<FieldName, ?> result = evaluator.evaluate(arguments); writeRecord(result, resultFields); ● The R script (data pre-and post-processing, model) is represented using standardized PMML data structures ● Well-defined data entry and exit interfaces
- 5. Workflow
- 6. R challenge Training: audit_train = read.csv("Audit.csv") audit_train$HourlyIncome = (audit_train$Income / (audit_train$Hours * 52)) audit.model = model(Adjusted ~ ., data = audit_train) saveRDS(audit.model, "Audit.rds") Deployment: audit_test = read.csv("Audit.csv") audit.model = readRDS("Audit.rds") # "Error in eval(expr, envir, enclos) : object 'HourlyIncome' not found" predict(audit.model, newdata = audit_test)
- 7. R solution Matrix interface ( ): label = audit$Adjusted features = audit[, !(colnames(audit) %in% c("Adjusted"))] # Returns a `model` object audit.model = model(x = features, y = label) Formula interface ( ): # Returns a `model.formula` object audit.model = model(Adjusted ~ . + I(Income / (Hours * 52)), data = audit)
- 8. From `pmml` to `r2pmml` library("pmml") library("r2pmml") audit = read.csv("Audit.csv") audit$Adjusted = as.factor(audit$Adjusted) audit.glm = glm(Adjusted ~ . + I(Income / (Hours * 52)), data = audit, family = "binomial") saveXML(pmml(audit.glm), "Audit.pmml") r2pmml(audit.glm, "Audit.pmml")
- 9. Package `pmml` ( ) <PMML> <DataDictionary> <DataField name="I(Income/(Hours * 52))" optype="continuous" dataType="double"/> </DataDictionary> <GeneralRegressionModel> <MiningSchema> <MiningField name="I(Income/(Hours * 52))"/> </MiningSchema> </GeneralRegressionModel> </PMML>
- 10. Package `r2pmml` ( ) <PMML> <DataDictionary> <DataField name="Income" optype="continuous" dataType="double"/> <DataField name="Hours" optype="continuous" dataType="double"/> </DataDictionary> <TransformationDictionary> <DerivedField name="Income/(Hours * 52)" optype="continuous" dataType="double"> <Apply function="/"><FieldRef field="Income"/><Apply function="*"><FieldRef field="Hours"/><Constant>52</Constant></Apply></Apply> </DerivedField> </TransformationDictionary> <GeneralRegressionModel> <MiningSchema> <MiningField name="Income"/> <MiningField name="Hours"/> </MiningSchema> </GeneralRegressionModel> </PMML>
- 11. Manual conversion (1/2) r2pmml::r2pmml r2pmml::decorate base::saveRDS RDS file JPMML-R library The R side The Java side
- 12. Manual conversion (2/2) The R side: audit.model = model(Adjusted ~ ., data = audit) # Decorate the model object with supporting information audit.model = r2pmml::decorate(audit.model, dataset = audit) saveRDS(audit.model, "Audit.rds") The Java side: $ git clone https://github.com/jpmml/jpmml-r.git $ cd jpmml-r; mvn clean package $ java -jar target/converter-executable-1.2-SNAPSHOT.jar --rds-input /path/to/Audit.rds --pmml-output /path/to/Audit.pmml
- 14. Sanitization (1/2) ?base::ifelse # Static replacement value as.formula(Adjusted ~ ifelse(is.na(Hours), 0, Hours)) # Dynamic replacement value as.formula(paste("Adjusted ~ ifelse((is.na(Hours) | Hours < 0 | Hours > 168),", mean(df$Hours), ", Hours)", sep = ""))
- 15. Sanitization (2/2) <DerivedField name="ifelse(is.na(Hours))" optype="continuous" dataType="double"> <Extension>ifelse(is.na(Hours), 0, Hours)</Extension> <Apply function="if"> <Apply function="isMissing"> <FieldRef field="Hours"/> </Apply> <Constant>0</Constant> <FieldRef field="Hours"/> </Apply> </DerivedField>
- 16. Binarization (1/2) ?base::ifelse as.formula(Adjusted ~ ifelse(Hours >= 40, "full-time", "part-time"))
- 17. Binarization (2/2) <DerivedField name="ifelse(Hours >= 40)" optype="categorical" dataType="string"> <Extension>ifelse(Hours >= 40, "full-time", "part-time")</Extension> <Apply function="if"> <Apply function="greaterOrEqual"> <FieldRef field="Hours"/> <Constant>40</Constant> </Apply> <Constant>full-time</Constant> <Constant>part-time</Constant> </Apply> </DerivedField>
- 18. Bucketization (1/2) ?base::cut # Static intervals as.formula(Adjusted ~ cut(Income, breaks = 3)) # Dynamic intervals as.formula(Adjusted ~ cut(log10(Income), breaks = quantile(log10(Income)))
- 19. Bucketization (2/2) <DerivedField name="cut(Income)" optype="categorical" dataType="string"> <Extension>cut(Income, breaks = 3)</Extension> <Discretize field="Income"> <DiscretizeBin binValue="(129,1.61e+05]"> <Interval closure="openClosed" leftMargin="129.0" rightMargin="161000.0"/> </DiscretizeBin> <DiscretizeBin binValue="(1.61e+05,3.21e+05]"> <Interval closure="openClosed" leftMargin="161000.0" rightMargin="321000.0"/> </DiscretizeBin> <DiscretizeBin binValue="(3.21e+05,4.82e+05]"> <Interval closure="openClosed" leftMargin="321000.0" rightMargin="482000.0"/> </DiscretizeBin> </Discretize> </DerivedField>
- 20. Mathematical expressions (1/2) ?base::I as.formula(Adjusted ~ I(floor(Income / (Hours * 52)))) Supported constructs: ● Arithmetic operators `+`, `-`, `*`, `/`, `^` and `**` ● Relational operators `==`, `!=`, `<`, `<=`, `>=` and `>` ● Logical operators `!`, `&` and `|` ● Functions `abs`, `ceiling`, `exp`, `floor`, `is.na`, `log`, `log10`, `round` and `sqrt`
- 21. Mathematical expressions (2/2) <DerivedField name="floor(Income/(Hours * 52))" optype="continuous" dataType="double"> <Extension>I(floor(Income/(Hours * 52)))</Extension> <Apply function="floor"> <Apply function="/"> <FieldRef field="Income"/> <Apply function="*"> <FieldRef field="Hours"/> <Constant>52</Constant> </Apply> </Apply> </Apply> </DerivedField>
- 22. Renaming and regrouping categories (1/2) library("plyr") ?plyr::revalue ?plyr::mapvalues as.formula(Adjusted ~ plyr::revalue(Employment, c("PSFederal" = "Public", "PSState" = "Public", "PSLocal" = "Public"))) as.formula(Adjusted ~ plyr::revalue(Education, c("Yr1t4" = "Yr1t6", "Yr5t6" = "Yr1t6", "Yr7t8" = "Yr7t9", "Yr9" = "Yr7t9", "Yr10" = "Yr10t12", "Yr11" = "Yr10t12", "Yr12" = "Yr10t12"))) as.formula(Adjusted ~ plyr::mapvalues(Gender, c("Male", "Female"), c(0, 1)))
- 23. Renaming and regrouping categories (2/2) <DerivedField name="revalue(Employment)" optype="categorical" dataType="string"> <Extension>plyr::revalue(Employment, c(PSFederal = "Public", PSState = "Public", PSLocal = "Public"))</Extension> <MapValues outputColumn="to"> <FieldColumnPair field="Employment" column="from"/> <InlineTable> <row><from>PSFederal</from><to>Public</to></row> <row><from>PSState</from><to>Public</to></row> <row><from>PSLocal</from><to>Public</to></row> <row><from>Consultant</from><to>Consultant</to></row> <row><from>Private</from><to>Private</to></row> <row><from>SelfEmp</from><to>SelfEmp</to></row> <row><from>Volunteer</from><to>Volunteer</to></row> </InlineTable> </MapValues> </DerivedField>
- 24. Interactions as.formula(Adjusted ~ # Continuous vs. continuous Age:Income + # Continuous vs. transformed continuous Age:I(log10(Income)) + # Continuous vs. categorical Gender:Income + # Categorical vs. categorical (Gender + Education) ^ 2 )
- 26. Alternative encodings (1/2) # Continuous label ~ Continuous and categorical features audit.glm = glm(Income ~ Age + Employment + Education + Marital + Occupation + Gender + Hours, data = audit, family = "gaussian") # Encoded as GeneralRegressionModel element r2pmml(audit.glm, "Audit.pmml") # Encoded as RegressionModel element r2pmml(audit.glm, "Audit.pmml", converter = "org.jpmml.rexp.LMConverter")
- 27. Alternative encodings (2/2) # Binary label ~ Categorical features audit.glm = glm(Adjusted ~ cut(Age, breaks = c(0, 21, 65, 100)) + Employment + Education + Marital + Occupation + cut(Income, breaks = quantile(Income)) + Gender + ifelse(Hours >= 40, "full-time", "part-time"), data = audit, family = "binomial") audit.scorecard = r2pmml::as.scorecard(audit.glm) # Encoded as Scorecard element r2pmml(audit.scorecard, "Audit.pmml")
- 29. Software (Nov 2017) ● The R side: ○ base 3.3(.1) ○ pmml 1.5(.2) ○ r2pmml 0.15(.0) ○ plyr 1.8(.4) ● The Java side: ○ JPMML-R 1.2(.20) ○ JPMML-Evaluator 1.3(.10)