EverywhereML

A Python package to train Machine Learning models that run (almost) everywhere, including:

• [X] C++ / embedded systems
• [X] Javascript
• [X] PHP
• [ ] Go / TinyGo
• [ ] MicroPython
• [ ] ... other languages

This means you can deploy your models to:

• Edge devices
• Web servers
• Web browsers
• ... other environments

Components

The package implements most of the tools you need to develop a fully functional model, including:

• [X] Data loading and visualization
• [X] Preprocessing
  • [ ] Pipeline
  • [X] BoxCox (power transform)
  • [X] CrossDiff
  • [X] MinMaxScaler
  • [X] Normalizer
  • [X] PolynomialFeatures
  • [X] RateLimit
  • [X] StandardScaler
  • [X] YeoJohnson (power transform)
• [ ] Audio
  • [ ] MelSpectrogram
• [X] Feature selection
  • [X] RFE
  • [X] SelectKBest
• [ ] Time series analysis
  • [X] Diff
  • [X] Fourier transform
  • [X] Rolling window
  • [ ] TSFRESH
• [ ] Classification
  • [X] RandomForest
  • [X] LogisticRegression
  • [X] GaussianNB
  • [ ] BernoulliNB
  • [ ] SVM (not tested)
  • [ ] LinearSVM
  • [X] DecisionTree
  • [X] XGBoost
  • [ ] Catboost
• [ ] Regression
• [ ] LinearRegression

Each of these components can be trained in Python and exported to any of the supported languages with no (or as few as possible) external dependencies.

For example:

from everywhereml.data.preprocessing import MinMaxScaler
from sklearn.datasets import load_iris

transformer = MinMaxScaler()
X, y = load_iris(return_X_y=True)
Xt, yt = transformer.fit_transform(X, y)

print('Original range', (X.min(), X.max()))
print('Transformed range', (Xt.min(), Xt.max()))

# port to C++
print(transformer.port(language='cpp'))

# port to Js
print(transformer.port(language='js'))

# port to PHP
print(transformer.port(language='php'))