LIBANN: A High-Performance Neural Network Library Written in Pure C

LIBANN is a pure ANSI-C neural network library maintained by mseminatore. Its core consists of only two files: tensor.c (tensor operations) and ann.c (training/inference runtime), balancing simplicity, portability, and ease of use. It supports cross-platform deployment (tested and verified on Windows, macOS, Linux, etc.), BLAS acceleration, and provides complete tensor operations, multiple optimizers, activation functions, and training/inference capabilities. It is suitable for embedded systems, teaching, rapid prototyping, and integration into existing C/C++ projects.

Original Author and Source

• Original Author/Maintainer: mseminatore
• Source Platform: GitHub
• Original Link: https://github.com/mseminatore/ann
• Release Date: 2026-06-07

Project Overview

LIBANN focuses on simplicity, portability, and ease of use, with only two core files to avoid dependency conflicts and deployment complexity. Written in pure ANSI-C, it can be compiled and run on almost any platform that supports C. The maintainer regularly tests it on Windows (x86/x64), macOS (Intel/M1), and Ubuntu Linux to ensure cross-platform compatibility.

Core Architecture Design

Following the "minimum viable" principle, it is divided into two modules:

1. Tensor Library (tensor.c): Provides basic vector/matrix operations (creation, destruction, copying, slicing, etc.), supports scalar/vectorized implementations, and can be accelerated by enabling USE_BLAS; covers operations like addition, multiplication, GEMM, etc., currently mainly supporting 2D tensors.
2. Neural Network Runtime (ann.c): Implements complete training/inference functions based on the tensor library.

Optimizers and Loss Functions

• Optimizers: Built-in SGD, Momentum, AdaGrad, RMSProp, Adam (officially recommended default) five algorithms.
• Loss Functions: Supports Mean Squared Error (MSE, for regression) and Categorical Cross-Entropy (for multi-class classification).

Activation Functions and Regularization

• Activation Functions: Provides seven types: no activation, Sigmoid, Softmax, ReLU, LeakyReLU, Tanh, Softsign.
• Regularization: Dropout (inverted dropout technique), L1/L2 regularization, gradient clipping to prevent overfitting and gradient explosion.

Learning Rate Scheduling Strategies

Built-in three schedulers:

1. Step Decay: Multiply by decay coefficient gamma every N epochs;
2. Exponential Decay: Multiply by gamma every epoch for smooth continuous decay;
3. Cosine Annealing: Decay from initial learning rate to minimum following a cosine curve; Supports custom scheduling functions.

Incremental Training and Online Learning

Supports preserving optimizer state, not resetting weights, single-sample training, and inference during training, suitable for streaming data, model fine-tuning, and online learning scenarios.

Automatic Hyperparameter Tuning

Implemented via the ann_hypertune module:

• Search Strategies: Grid search, random search, Bayesian optimization, TPE;
• Tunable Parameters: Learning rate, batch size, optimizer type, number of hidden layers/neurons, network topology, activation functions, etc.

Model Export and Visualization

• ONNX Format: Export to ONNX JSON for easy cross-toolchain integration;
• PIKCHR Diagrams: Export network architecture as SVG vector graphics (small networks show node connections, large networks show layer block diagrams);
• Learning Curves: Export training history as CSV, including loss values and learning rates, which can be used to draw analysis charts.

BLAS Acceleration

Enabling BLAS can significantly improve training performance, supporting CBLAS (recommended for cross-platform) and OpenBLAS; enable it via CMake configuration (e.g., -DUSE_CBLAS=1), and call cblas_init() to initialize when using CBLAS.

Applicable Scenarios

• Embedded Systems: Pure C code, minimal dependencies, easy to port;
• Teaching Purposes: Clean code, suitable for learning neural network principles;
• Rapid Prototyping: Verify ideas without complex environments;
• Existing Project Integration: Add neural network capabilities to C/C++ projects by including just two files.

Summary

LIBANN implements a fully functional neural network library with a minimalist architecture, zero dependencies, and strong portability, providing an excellent choice for developers seeking simplicity, control, and portability.