Generating Scoring Code from the Formula Depot Platform
Scoring code generation is intended to facilitate using models built in JMP in a production environment or other framework of your choice. Many platforms publish prediction formulas to the Formula Depot, however; not all prediction formulas generate complete code. Code can be complete, it can be a code fragment, or the code can include unsupported functions that require additional programming for implementation. Error messages often indicate an unsupported function call that requires additional programming in the targeted language. Target language specific libraries might be required for the implementation of more complex functions. These libraries are discussed in the language specific sections below. For example, you can perform the following tasks:
• Deploy your model to SAS Model Manager using the generated SAS code.
• Augment your ETL process with in-database scoring using the generated SQL code.
• Create a node for a data transformation pipeline with an application built with the generated C code.
• Create a Jupyter notebook to show live scoring results using the generated Python code.
• Enable customers to score their own data with a web application that includes the generated JavaScript.
For the C, Python, and JavaScript languages, you must include supporting code such as .h files and utility libraries when deploying or compiling the generated code. These files are available in your JMP installation folder inside the Scoring folder.
Tip: To find the location of installation folder on your machine, use the JSL command Get Path Variable("$ALL_HOME/Scoring");.
C Code
The C scoring code that is generated must be compiled into a library and then linked into an application. You might use either a static or a dynamic link approach. The files jmp_lib.h, jmp_parms.h, and jmp_score.h needed for compiling and linking can be found in the Scoring/C folder in your JMP installation folder.
Python Code
The file jmp_score.py that is needed to run your Python scoring application can be found in the Scoring/Python folder in your JMP installation folder. When deploying your models, copy jmp_score.py to the same location as your executable scoring code.
When calling the generated code to score your data, it is important that the input data is passed in a data structure compatible with the code generated by JMP. The Python scoring code expects the input and output arguments to be of a dictionary-like type. Values can be accessed and written by key, such as DataFrame row objects from the Spark, Dask and Pandas frameworks, or the standard Python “dict” class.
Use of linear algebra operators or advanced operators such as Vec Quadratic and Design Norm require the NumPy Python library.
JavaScript Code
The file jmp_score.js that is needed to run your JavaScript scoring application can be found in Scoring/JavaScript folder.
SAS Code
The generated code fragment that once wrapped by PROC DS2 statements can be used in SAS applications including the SAS In-Database Code Accelerator. The code is added to a SAS Window and includes a variable name mapping section in the comment block that precedes the code. Any temporary variable names associated with intermediate columns are deleted at the end of the DATA Step by the drop command.
Tip: For models with an ifmax call, such as logistic or neural models for a categorical response, move temporary variable declarations before the method run() statement.
SQL Code
The SQL code fragment that once wrapped in a select statement can be used in SQL queries against most major database servers.
Note: Code that contains “placeholder” or “ERROR” indicates an unsupported function call.