nubnet-rs

This repository contains a simple (newbie) implementation an MLP in Rust built from scratch with minimum dependencies (only rand crate for the library code). Originally, this project emerged as a requirement for the PV021 course at FI MUNI. The task involved coding a neural network from the ground up to achieve an 88% accuracy on the Fashion MNIST dataset, a goal that has been successfully met.

The code has been made modular so that the network library can be used for other tasks as well.

Project structure

• src/ contains the library code
• examples/ contains code of how to use the library (e.g. training on MNIST)

Running Fashion MNIST example

1. Download the fashion MNIST dataset from here and extract it at the root of the repository (you should have data folder next to src and examples).
2. Run the training with cargo run --example mnist --release

Remarks

• It seems that the functional programming approach is not the best fit for a lot of immutable data structures, especially when they are nested.
• The primary emphasis was on achieving speed. Notably, a single epoch of training on the Fashion MNIST dataset takes approximately 2 seconds on an M1 MacBook Air.
• Multi-threading helps a lot, but strangely only up to 32 threads, after which it gets slower again (probably due to communication overhead).