patryklaskowski/PMML_Project

patryk.laskowski@ibm.com

---

1a. Python PMML supporting libraries

A. lightpmmlpredictor

But supports only PMML-4_1,
and we have PMML-4_3.

❌ REJECTED

B. Augustus - open-source PMML toolkit written in Python

GitHub page.
But hard to setup.

❌ REJECTED

C. openscoring-python

But there is no free license.

❌ REJECTED

D. pmml2sklearn

But parse pmml files and convert it to sklearn kmeans models.

❌ REJECTED

E. sklearn-pmml-model

But supports only few models (Decision Trees, Random Forests (ensemble method), Linear Model (specifically: Linear Reg, Ridge, Lasso, ElasticNet), Gaussian Naive Bayes)
Almost good...

❌ REJECTED

F. PyPmml

Which is simple, supports almost all models on all PMML versions and works!

✅ ACCEPTED

---

1b. R PMML supporting libraries

A. r-pmml

GitHub Supported models
Supports wide range of different model types (Anomaly Detection, Clustering, K Nearest Neighbors, Linear Models, etc.)

But this is PMML reader for R language.

This package contains functions to export various machine learning and statistical models to PMML, as well as generate data transformations in PMML format.

❌ REJECTED

B. r2pmml

R package for converting R models to PMML. Requires Java.

❌ REJECTED

---

1c. Java PMML supporting libraries

A. jpmml-evaluator

Java Evaluator API for Predictive Model Markup Language (PMML)

Supported versions: PMML versions 3.0, 3.1, 3.2, 4.0, 4.1, 4.2, 4.3 and 4.4.
Wide range of models supported (Association rules, Cluster model, General regression, Naive Bayes, k-Nearest neighbors, Neural network, etc.)

✅ ACCEPTED

---

---

2. Python PyPmml [main solution]

What is PyPmml?

autodeployai/pypmml

PyPMML is a Python PMML scoring library, it really is the Python API for PMML4S.

What is PMML4S?

autodeployai/pmml4s

PMML4S is a PMML (Predictive Model Markup Language) scoring library for Scala. It provides both Scala and Java Evaluator API for PMML. PMML4S is a lightweight, clean and efficient implementation based on the PMML specification from 2.0 through to the latest 4.4.

Example PyPmml code.

Below code load PMML decision tree model and make prediction.

>>> from pypmml import Model

>>> model = Model.load('Waga.xml')
>>> model.predict({'Wzrost': 180})
{'$RI-Waga': 24, '$R-Waga': 65.32}
>>> model.inputNames
['Wzrost']
>>> model.predict([180])
[65.32, 24]
>>> model.outputNames
['$R-Waga', '$RI-Waga']

Data in Pandas DataFrame form is also accepted.

>>> import pandas as pd

>>> data = pd.DataFrame({
...     'Wzrost' : [180, 170, 160]})

>>> model.predict(data)
     $R-Waga  $RI-Waga
0  65.320000        24
1  57.190909        40
2  51.010526        17

---

3. Support for model types

Note

In this case PMML models are going to be produced in SAS.
Below compare all possible models that PMML supports with all models that SAS may export in PMML format with all models that PyPmml Python library as well as JPmml support.

---

Idx	All PMML models	SAS Enterprise Miner [PMML Producer]	SAS Enterprise Miner 14.3 [PMML Consumer/Producer]	PyPMML [Python Library]	JPMML-Evaluator [Java Evaluator]
1	AnomalyDetectionModel	-	-	✅ Anomaly Detection Models	-
2	AssociationModel	✅ Association Rules	-	✅ Association Rules	✅ Association rules
3	BayesianNetworkModel	-	-	❌ Bayesian Network	-
4	BaselineModel	-	✅ Baseline models [???]	❌ Baseline Models	-
5	ClusteringModel	✅ Clustering	✅ Clustering models	✅ Cluster Models	✅ Cluster model
6	GaussianProcessModel	-	-	❌ Gaussian Process	-
7	GeneralRegressionModel	-	-	✅ General Regression	✅ General regression
8	MiningModel	-	-	-
9	NaiveBayesModel	-	✅ Naïve Bayes [PMML Consumer]	✅ Naive Bayes	✅ Naive Bayes
10	NearestNeighborModel	-	-	✅ k-Nearest Neighbors	✅ k-Nearest neighbors
11	NeuralNetwork	✅ Neural Networks	✅ Neural Networks	✅ Neural Network	✅ Neural Networks
12	RegressionModel	✅ Linear Regression ✅ Logistic Regression	✅ Regression	✅ Regression	✅ Regression
13	RuleSetModel	-	-	✅ Ruleset	✅ Rule set
14	SequenceModel	-	-	❌ Sequences	-
15	Scorecard	-	✅ Scorecard [PMML Consumer]	✅ Scorecard	✅ Scorecard
16	SupportVectorMachineModel	-	✅ Vector Machine [PMML Consumer]	✅ Vector Machine	✅ Support Vector Machine
17	TextModel	-	-	❌ Text Models	-
18	TimeSeriesModel	-	-	❌ Time Series	✅ Time Series
19	TreeModel	✅ Decision Trees	✅ Trees	✅ Trees	✅ Tree Model

---

3a. All PMML models

Documentation points these models as supported by PMML format.

---

3b. SAS Enterprise Miner

According to this website SAS Enterprise Miner can produce this models:

---

3c. SAS Enterprise Miner 14.3 PMML Support

Website. Look for "Supported PMML Models".

SAS PROC PSCORE supports the following types of PMML models:

This means that SAS in general is able to handle either import or export of above.

---

3d. Python PyPmml

Supports following models:

---

3e. Java jpmml-evaluator

Evaluate:

---

Summary

PyPmml is a good solution that supports wide range of PMML versions and all desired models except Baseline Models. This seems to be a great choice for this project!

---

patryk.laskowski@ibm.com