Kianqunki / sklearn-porter

Transpile trained scikit-learn estimators to C, Java, JavaScript and others.

sklearn-porter

Transpile trained scikit-learn estimators to C, Java, JavaScript and others.
It's recommended for limited embedded systems and critical applications where performance matters most.

Estimators

Estimator	Programming language
Classifier	Java *	JS	C	Go	PHP	Ruby
svm.SVC	✓, ✓ ᴵ	✓	✓		✓	✓
svm.NuSVC	✓, ✓ ᴵ	✓	✓		✓	✓
svm.LinearSVC	✓, ✓ ᴵ	✓	✓	✓	✓	✓
tree.DecisionTreeClassifier	✓, ✓ ᴱ, ✓ ᴵ	✓, ✓ ᴱ	✓, ✓ ᴱ	✓, ✓ ᴱ	✓, ✓ ᴱ	✓, ✓ ᴱ
ensemble.RandomForestClassifier	✓ ᴱ, ✓ ᴵ	✓ ᴱ	✓ ᴱ	✓ ᴱ	✓ ᴱ	✓ ᴱ
ensemble.ExtraTreesClassifier	✓ ᴱ, ✓ ᴵ	✓ ᴱ	✓ ᴱ		✓ ᴱ	✓ ᴱ
ensemble.AdaBoostClassifier	✓ ᴱ, ✓ ᴵ	✓ ᴱ, ✓ ᴵ	✓ ᴱ
neighbors.KNeighborsClassifier	✓, ✓ ᴵ	✓, ✓ ᴵ
naive_bayes.GaussianNB	✓, ✓ ᴵ	✓
naive_bayes.BernoulliNB	✓, ✓ ᴵ	✓
neural_network.MLPClassifier	✓, ✓ ᴵ	✓, ✓ ᴵ
Regressor	Java *	JS	C	Go	PHP	Ruby
neural_network.MLPRegressor		✓

✓ = is full-featured, ᴱ = with embedded model data, ᴵ = with imported model data, * = default language

Installation

Stable

$ pip install sklearn-porter

Development

If you want the latest changes, you can install this package from the master branch:

$ pip uninstall -y sklearn-porter
$ pip install --no-cache-dir https://github.com/nok/sklearn-porter/zipball/master

Usage

Export

The following example demonstrates how you can transpile a decision tree estimator to Java:

from sklearn.datasets import load_iris
from sklearn.tree import tree
from sklearn_porter import Porter

# Load data and train the classifier:
samples = load_iris()
X, y = samples.data, samples.target
clf = tree.DecisionTreeClassifier()
clf.fit(X, y)

# Export:
porter = Porter(clf, language='java')
output = porter.export(embed_data=True)
print(output)

The exported result matches the official human-readable version of the decision tree.

Integrity

You should always check and compute the integrity between the original and the transpiled estimator:

# ...
porter = Porter(clf, language='java')

# Compute integrity score:
integrity = porter.integrity_score(X)
print(integrity)  # 1.0

Prediction

You can compute the prediction(s) in the target programming language:

# ...
porter = Porter(clf, language='java')

# Prediction(s):
Y_java = porter.predict(X)
y_java = porter.predict(X[0])
y_java = porter.predict([1., 2., 3., 4.])

Notebooks

You can run and test all notebooks by starting a Jupyter notebook server locally:

$ make open.examples
$ make stop.examples

CLI

In general you can use the porter on the command line:

$ porter <pickle_file> [--to <directory>]
         [--class_name <class_name>] [--method_name <method_name>]
         [--export] [--checksum] [--data] [--pipe]
         [--c] [--java] [--js] [--go] [--php] [--ruby]
         [--version] [--help]

The following example shows how you can save a trained estimator to the pickle format:

# ...

# Extract estimator:
joblib.dump(clf, 'estimator.pkl', compress=0)

After that the estimator can be transpiled to JavaScript by using the following command:

$ porter estimator.pkl --js

The target programming language is changeable on the fly:

$ porter estimator.pkl --c
$ porter estimator.pkl --java
$ porter estimator.pkl --php
$ porter estimator.pkl --java
$ porter estimator.pkl --ruby

For further processing the argument --pipe can be used to pass the result:

$ porter estimator.pkl --js --pipe > estimator.js

For instance the result can be minified by using UglifyJS:

$ porter estimator.pkl --js --pipe | uglifyjs --compress -o estimator.min.js

Development

Environment

You have to install required modules for broader development:

$ make install.environment  # conda environment (optional)
$ make install.requirements.development  # pip requirements

Independently, the following compilers and intepreters are required to cover all tests:

Name	Version	Command
GCC	>=4.2	gcc --version
Java	>=1.6	java -version
PHP	>=5.6	php --version
Ruby	>=2.4.1	ruby --version
Go	>=1.7.4	go version
Node.js	>=6	node --version

Testing

The tests cover module functions as well as matching predictions of transpiled estimators. Start all tests with:

$ make test

The test files have a specific pattern: '[Algorithm][Language]Test.py':

$ pytest tests -v -o python_files='RandomForest*Test.py'
$ pytest tests -v -o python_files='*JavaTest.py'

While you are developing new features or fixes, you can reduce the test duration by changing the number of tests:

$ N_RANDOM_FEATURE_SETS=5 N_EXISTING_FEATURE_SETS=10 \
  pytest tests -v -o python_files='*JavaTest.py'

Quality

It's highly recommended to ensure the code quality. For that Pylint is used. Start the linter with:

$ make lint

Citation

If you use this implementation in you work, please add a reference/citation to the paper. You can use the following BibTeX entry:

@unpublished{skpodamo,
  author = {Darius Morawiec},
  title = {sklearn-porter},
  note = {Transpile trained scikit-learn estimators to C, Java, JavaScript and others},
  url = {https://github.com/nok/sklearn-porter}
}

License

The module is Open Source Software released under the MIT license.

Questions?

Don't be shy and feel free to contact me on Twitter or Gitter.