MiniNeuron

train feedforward neural networks in C++ — only dependency is optional OpenMP

build:

git clone https://github.com/ligumas/MiniNeuron
cd MiniNeuron
cmake -B build && cmake --build build

Quick example (XOR)

#include "MiniNeuron/includes/network.h"

MiniNeuron::Network net;

// structure: 2 → 4 → 1
net.add(MiniNeuron::Layer(4, 2, ActivationType::Sigmoid, InitializerType::Xavier));
net.add(MiniNeuron::Layer(1, 4, ActivationType::Sigmoid, InitializerType::Xavier));
net.initLayers();

// inputs and targets are std::vector<std::vector<float>>
net.train(inputs, targets, 1, 500, 5.0f, LossTypes::crossEntropy);

std::vector<float> prediction = net.forward(input);

// save to a .mn file
net.saveModel("xor-model.mn");

Features

• Dense (fully connected) layers
• Activations: Sigmoid, ReLU, Tanh, LeakyReLU, Softmax, Linear
• Loss functions: MSE, Cross Entropy
• Optimizers: SGD, Adam
• Initializers: Xavier, He
• Multi-threading via OpenMP (optional)
• Model save/load (.mn format)

Benchmarks

v1.1.0 — with OpenMP

Demo	Network	Epochs	Accuracy	Time	CPU
XOR	2→4→1	500	Loss: 0.0133	~0s	AMD Ryzen 5 5600
MNIST (Sigmoid)	784→128→10	1	95.65%	~1s	AMD Ryzen 5 5600
MNIST (ReLU)	784→256→128→10	3	97.36%	~12s	AMD Ryzen 5 5600

v1.0.0 — single-threaded

Demo	Network	Epochs	Accuracy	Time	CPU
XOR	2→4→1	500	Loss: 0.01	~1s	AMD Ryzen 5 5600
MNIST (Sigmoid)	784→128→10	1	95.66%	~5min	AMD Ryzen 5 5600
MNIST (ReLU)	784→256→128→10	3	96.98%	~37min	AMD Ryzen 5 5600

License

MIT