cpp_deep_neural_net

Neural networks from scratch in modern C++17 — zero third‑party dependencies.

---

✨ Features

• Pure C++17 implementation — no Eigen, no BLAS, no external frameworks
• Xavier & Kaiming weight initialisation
• Mini‑batch SGD (with momentum) & Adam optimiser
• Runs the full MNIST pipeline end‑to‑end
• Single‑command build with Make

Work‑in‑progress: convolutional layers, residual connections & transformer blocks.

---

🗂️ Repository layout

.
├── src/            # C++ source 
│   ├── neural_net.h    # Implementation of the Neural Network
│   ├── inference.cpp   # Run inference on trained weights
│   └── main.cpp        # Train and save model weights
├── scripts/        # Helper Python scripts (dataset download)
|   └── fetch_mnist.py
├── params/         # Hyper‑parameter files
├── weights/        # Saved checkpoints 
├── logs/           # Training logs 
└── README.md

---

🚀 Quick start

Prerequisites

• Linux or macOS
• g++ ≥ 10 (C++17)
• GNU make
• Python 3 (optional, for helper scripts)

Build

Firstly, perform the following:

# Clone
$ git clone https://github.com/GunaDD/cpp_deep_neural_net.git
$ cd cpp_deep_neural_net

How to prepare the dataset:

Download the MNIST data and place it at data/mnist. It should contain:

t10k-images-idx3-ubyte
t10k-labels-idx1-ubyte
train-images-idx3-ubyte
train-labels-idx1-ubyte

If the filename of thes data differ, make sure to adjust it to this.

Then run pip3 install numpy if you don't already have numpy.

Then run make data from the cpp_deep_neural_det directory. You should expect to see (on the data/mnist directory) :

t10k-images.f32
t10k-labels.u8
train-images.f32
train-labels.u8

How to train the model:

make main
touch ./logs/train_[experiment_name]
./bin/main > ./logs/train_[experiment_name]

Before running the commands, you can tune the model parameters (e.g. num of layers, learning rate schedule, weight initializations, experiment name) from the main.cpp file

How to inference:

make inference
./bin/inference > ./logs/inference_[experiment_name]

Before running the command, you can specify the experiment name and the epoch number of the weights you want to load for to be used during inference time.

Executables are placed in bin/.

---

📊 Results (MNIST)

Model	Parameters	Test accuracy
full_trainset_200_epoch	26506	92.91%

---

📝 License

This project is licensed under the MIT License — see the LICENSE file for details.