ppmml

ppmml is a python library for converting machine learning models to pmml file. ppmml wraps jpmml libraries and provides clean interface.

Installation

pip install --default-timeout=10000 -i https://pypi.anaconda.org/lgrcyanny/simple ppmml

If download too slow, please download from anaconda in ppmml conda package, then run the command:

pip install ppmml-0.0.1.tar.gz

Geting Started

import pandas as pd
from sklearn.datasets import load_iris
from sklearn.linear_model import LogisticRegression
from sklearn.externals import joblib
import ppmml
# load data and train iris datasets
(X, y) = load_iris(True)
lr = LogisticRegression(tol=1e-5)
lr.fit(X, y)
joblib.dump(lr, "lr.pkl.z", compress = 9)

# to pmml file
ppmml.to_pmml("lr.pkl.z", "lr.pmml", model_type='sklearn')

# prepare test data
df = pd.DataFrame(X)
df.columns = ['x1', 'x2', 'x3', 'x4']
df.to_csv('test.csv', header=True, index=False)
# predit with pmml file, a simple predict API based on jpmml-evaluator
ppmml.predict('lr.pmml', 'test.csv', 'predict.csv')

More examples

• sklearn models to pmml file
• xgboost model to pmml file
• lightgbm model to pmml file
• tensorflow model to pmml file
• r model to pmml file

Algorithm Features

All algorithm supported by jpmml is support by pmml. In summary:

sklearn estimators

Supervised Learning

• GLM: Linear, Logistic Regression, Lasso, ElasticNet, Ridge, SGD
• Naive Bayes: GaussianNB, but no support for multinomial naive Bayes, Bernoulli naive Bayes
• Nearest Neighbors
• Neural Network
• SVM
• DecisionTree
• All EM Method
• LinearDiscriminantAnalysis

Unsupervised Learning

• Custering: KMeans, no support for LDA and DBSCAN
• PCA

feature algorithms

• feature selection
• feature extraction, no support for FeatureHasher
• feature selection
• feature preprocessing

jpmml-sklearn

xgboost classifier and regressor

jpmml-xgboost

lightgbm classifier and regressor

jpmml-lightgbm

tensorflow

Only support DNNClassifier, DNNRegressor, Linear Classifier, Linear Regressor, one_hot_column, real_valued_column sparse_column_with_keys jpmml-tensorflow

spark ml

jpmml-sparkml is better than spark mllib's pmml transformation. it support 35 algorithms now.

• Feature extractors, transformers and selectors. But no support for BucketedRandomProjectionLSH,DCT, ElementWiseProduct, LSH, MinHashLSH, Normalizer, PolynomialExpansion
• Classification: LR, GBT, DecisionTree, NN, RandomForest. But no support for SVM，Naive Bayes.
• Regression: Linear, GBT, DecisionTree, RandomForest, GLR. But no support for Survival regression, Isotonic regression
• Clustering: KMeans. But no support for GMM, LDA

R models

ppmml integrates jpmml-r

Preprocessing

• range, center, scale, medianImpute

R Algorithms

• glm - Generalized linear (GLM) regression and classification
• kmeans - K-Means clustering
• lm - Linear (LM) regression
• XGBoost, GBM
• Random Forest
• SVM Classifier and Regression
• Scorecard regression
• earth - Multivariate Adaptive Regression Spline (MARS) regression
• elmNN - Extreme Learning Machine (ELM) regression
• iForest - Isolation Forest (IF) anomaly detection
• mvr - Multivariate Regression (MVR) regression
• lrm - Binary Logistic Regression (LR) classification
• ols - Ordinary Least Squares (OLS) regression

Requirements

• python 2.7
• jdk 1.8+

How to build from source

sh build.sh clean package

the output egg package will be palced in dist directory

install to local

python setup.py install

Project Structure

• ppmml: ppmml python libraries
• deps: jar dependencies, including jpmml-sklearn, jpmml-tensorflow, jpmml-r, jpmml-spark, jpmml-xgboost, jpmml-lightgbm and jpmml-evaluator.
• examples: ppmml example notebooks

Run unit tests

please refer to dev guide

All unit tests are passed with these versions:

• tensorflow 1.4
• xgboost 0.6a2
• scikit-learn 0.19
• lightgbm 2.0.11
• spark 2.2, 2.3
• R 3.4.2
• jpmml-model 1.3.8

Notes

Notes for Users

• pmml converters only support run in locally, especially spark converter will new a local SparkSession
• Users care about the input path and pmml output path

Notes for developers

• jpmml-tensorflow hasn't been publish to maven, please pull the code and compile the jar manually (https://github.com/lgrcyanny/jpmml-tensorflow/tree/fix-building) The forked version has fixed compile error of jpmml-tensorflow