tinycml

Overview

tinycml is a minimal, pure C11 machine learning library with zero external dependencies. Neural networks, grid search, preprocessing — all under 12K LOC.

Designed for resource-constrained environments where pulling in Python or C++ dependencies isn’t an option.

Features

• Neural Networks — Feedforward networks with configurable layers, activations (sigmoid, tanh, ReLU), and training via backpropagation
• Grid Search — Generic k-fold cross-validation with hyperparameter tuning (qsort-based)
• Preprocessing — Normalization, standardization, train-test split
• Metrics — Accuracy, precision, recall, F1-score, confusion matrix
• Zero Dependencies — Pure C11, no external libs, no dynamic allocation leaks
• Thread-safe — No global state, all state passed via struct pointers

Architecture

tinycml/
├── cml_matrix.h      — Matrix operations (transpose, multiply, inverse)
├── cml_nn.h          — Neural network (forward, backprop, train, predict)
├── cml_preprocess.h  — Normalization, standardization, splitting
├── cml_metrics.h     — Classification metrics
├── cml_grid.h        — Grid search with k-fold CV
├── cml.h             — Umbrella include header
└── examples/         — Usage examples

Status

Active development. Phase 1 (core API, NN, grid search) is complete. Phase 2 (tests + CI) and Phase 3 (performance optimizations) are in progress.