mnist

Zig implementation of a simple neural network for mnist

Getting started

First, follow the instructions at https://ziglang.org/download. I'm using version 0.13.0.

Then, download the dataset from http://yann.lecun.com/exdb/mnist/. The four files you need are:

train-images-idx3-ubyte: training set images
train-labels-idx1-ubyte: training set labels
t10k-images-idx3-ubyte:  test set images
t10k-labels-idx1-ubyte:  test set labels

Without renaming the files, put them in a data/ folder at the root of this project's directory.

zig build run

To actually make it go fast:

zig build -Doptimize=ReleaseFast run

On my machine, getting around 93% accuracy with 100 epochs.

To run tests:

zig build test

Motivation

This project was a way for me to (1) learn the fundamentals of neural networks and (2) learn Zig.

I started with Michael Nielsen's excellent introduction to neural networks and the Python implementation within as a reference. But translating the Python code directly to Zig was a clumsy effort. By chance, I stumbled upon this blog post implementing mnist in Zig. Taking inspiration from there, I refactored my original implementation to hard-code the layers and feedforward/backprop them explicitly instead of using a loop.