tiny-technologies / neural

A simple neural network implement in C

Neural

This repo contains a simple implementation of a neural network + backpropagation written in C.

Installation

You can run neural using Nix:

nix run github:tiny-technologies/neural

Or, install it:

nix profile install github:tiny-technologies/neural

Usage

Here is a summary of the available commands and their respective arguments:

Run

Predict the label of an image using a trained network:

neural run <path_to_model> <path_to_image>

The image needs to be in the PGM P5 format. You can edit the example.pgm with an image manipulation tool of your choice (e.g. GIMP).

Test

Test the accuracy of a trained network on the test data:

neural test <path_to_model>

Train

To finetune an existing model or train a new one from scratch, use the

neural train

subcommand. For detailed list of command line flags see below.

Help

Show all subcommands and their respective options:

neural help

Usage:

    neural <command> [<args>]

Commands:

    run    Run inference using a trained network
      <path>                      path to model
      <path>                      path to PGM P5 image 

    test   Test the accurary of a trained network
      <path>                      path to model (default: default.model)

    train  Train a new network and store it to disk
      -b, --batch-size <int>      samples per batch (default: 200)
      -d, --dims <int,int,..>     dimensions of hidden layers (default: 16,16)
      -e, --epochs <int>          number of epochs (default: 10)
      -l, --learning-rate <real>  step size of parameter update (default: 0.01)
      -i, --input <path>          path to model used as starting point (optional)
      -o, --output <path>         output path of the trained model (default: default.model)

    bench  Benchmark forward and backward pass

    help   Show this message and exit

Development

Tooling

When you cd into this repository, direnv activates a Nix-environment containing all required development tools.

Workflow

The just command is the main entry point into this repository. It is a convenient way to execute commands and has a documenting purpose by making important project-specific commands visible.

To list all commands, run:

just

Download the MNIST dataset:

just download

Compile and run the debug build:

just debug

Compile and run the release build:

just release

Compile and run tests:

just test

Compile in release mode, with optimizations:

just build release

License

GNU General Public License v3.0