A from-scratch implementation of classic machine learning algorithms — no scikit-learn under the hood, just NumPy. Built for understanding how these models actually work: the math, the gradients, the splits.
Closed-form (normal equations) and gradient descent solvers, with optional L1/L2 regularization.
Gradient descent with a sigmoid decision function, optional L1/L2 regularization.
Gini-impurity splitting, recursive binary partitioning, configurable max depth.
Instance-based classification (and regression) via majority vote / mean over the k closest training points, vectorized distance computation.
Plots were generated with python docs/generate_plots.py — rerun it after changing a model to refresh them.
glassboxml/ Core package: models, losses, optimizers, metrics, data generators
tests/ pytest suite covering the package
experiments/ Exploratory scripts and notebooks, one folder per algorithm
webapp/ FastAPI + Chart.js demo - train a model and see it plotted in the browser
scripts/ build.sh / test.sh / run.sh - build, test, and run the Dockerized webapp
docs/ README images and the script that generates them
k8s/ Deployment/Service manifests for running the webapp on minikube
Dockerfile Container image for the webapp
Jenkinsfile Declarative pipeline: checkout -> install -> test -> docker build
.github/workflows/ GitHub Actions: test suite + Docker build on push/PR to main
Requires Python 3.11+.
pip install -e .import numpy as np
from glassboxml.models.linear_regression import LinearRegression
from glassboxml.data.generators import generate_regression_dataset
X, y, w_true, b_true = generate_regression_dataset(w_true=[2.5], b_true=1.0, n_samples=200)
model = LinearRegression()
model.fit_gradient_descent(X, y, epochs=500, learning_rate=0.1)
print(model.w, model.b)More end-to-end examples, including regularization sweeps and decision-boundary plots, live in experiments/.
A small FastAPI app lets you tweak parameters (sample size, noise, learning rate, tree depth) and watch each model fit live in the browser via Chart.js.
pip install -e . && pip install -r webapp/requirements.txt
uvicorn webapp.app:app --reloadThen open http://localhost:8000.
pytest
# or
./scripts/test.sh./scripts/build.sh # docker build -t glassboxml-webapp .
./scripts/run.sh # docker run -p 8000:8000 glassboxml-webapp- GitHub Actions (.github/workflows/tests.yml) runs the test suite and validates the Docker build on every push/PR to
main. - Jenkinsfile implements the same pipeline (checkout, install, test, Docker build) as a declarative Jenkins pipeline, for running the same checks outside GitHub Actions.
k8s/deployment.yaml and k8s/service.yaml deploy the containerized webapp to a local minikube cluster (NodePort service, since it's local-only). See the manifests for the exact image-loading step required before kubectl apply.
MIT — see LICENSE.