tinysvm 0.1.0

A lightweight, dependency-free (NumPy only) SVM implementation.

TinySVM 🚀

A lightweight, dependency-free (NumPy only), and educational implementation of Support Vector Machines (SVM) in pure Python.

TinySVM implements both Classification (SVC) and Regression (SVR) with a simplified SMO (Sequential Minimal Optimization) and Coordinate Descent algorithm. It serves as a great learning resource or a drop-in replacement for heavy libraries when you only need basic SVM functionality.

✨ Features

• Zero Dependencies: Only requires numpy. No scikit-learn or scipy needed.
• Full-featured:
  • Binary & Multi-class Classification (One-vs-Rest).
  • Regression (Single & Multi-output).
  • Kernel Support: Linear & RBF (Gaussian).
• Production Ready-ish:
  • Built-in Auto-Scaling (StandardScaler).
  • Probability estimates (predict_proba).
  • Scikit-learn compatible API (fit, predict, score).
• Tiny: Less than 300 lines of code.

📦 Installation

Just copy tinysvm.py to your project folder. Yes, it's that simple!

Dependencies:

pip install numpy

⚡ Quick Start

Classification (XOR Problem)

from tinysvm import TinySVM
import numpy as np

# XOR data (Linear Separable? No.)
X = [[0, 0], [1, 1], [1, 0], [0, 1]]
y = [0, 0, 1, 1]

# Initialize with RBF kernel and Auto-Scaling
clf = TinySVM(mode='classification', kernel='rbf', gamma=2.0, C=10.0, scaling=True)
clf.fit(X, y)

print(f"Prediction: {clf.predict([[0, 1]])}") # Output: [1]
print(f"Accuracy:   {clf.score(X, y)}")       # Output: 1.0

Regression

# Simple Linear Regression
X = [[1], [2], [3], [4], [5]]
y = [3, 5, 7, 9, 11] # y = 2x + 1

reg = TinySVM(mode='regression', kernel='linear', C=50.0)
reg.fit(X, y)

print(f"Prediction for x=6: {reg.predict([[6]])}") # Should be close to 13

⚙️ API Reference

TinySVM(mode, C, kernel, gamma, scaling, ...)

• mode: 'classification' or 'regression'.
• C: Regularization parameter (default 1.0).
• kernel: 'rbf' (default) or 'linear'.
• gamma: Kernel coefficient for RBF.
• scaling: Boolean. If True (default), automatically scales data using Z-score.

Methods

• fit(X, y): Train the model.
• predict(X): Predict class or value.
• predict_proba(X): (Classification only) Estimate class probabilities.
• score(X, y): Returns Accuracy (Classification) or R² (Regression).
• save(path) / load(path): Save/Load model state.

📜 License

MIT License. Feel free to use it in your own projects!