An experiment in following neuralnetworksanddeeplearning.com in Rust.

Go to file

q3k 61dbe86c2b prepare stuff for public code dump		2023-04-02 13:55:20 +02:00
bigbrain	js split	2021-04-03 13:05:13 +02:00
bigbrainjs	js split	2021-04-03 13:05:13 +02:00
proto	export to proto	2021-04-03 00:21:50 +02:00
src	prepare stuff for public code dump	2023-04-02 13:55:20 +02:00
.gitignore	feedforward	2021-03-28 15:07:19 +02:00
COPYING	prepare stuff for public code dump	2023-04-02 13:55:20 +02:00
Cargo.lock	js split	2021-04-03 13:05:13 +02:00
Cargo.toml	js split	2021-04-03 13:05:13 +02:00
README.md	prepare stuff for public code dump	2023-04-02 13:55:20 +02:00
shell.nix	feedforward	2021-03-28 15:07:19 +02:00

README.md

bigbrain

An experiment in following neuralnetworksanddeeplearning.com in Rust. The end result is a web page that can sometimes recognize hand-written digits using a three-layer (one hidden) neural network.

You can see it in action at https://hackerspace.pl/~q3k/bigbrain/main.html. It's really good at figuring out the digit 2, and gets somewhat confused about other digits. Be kind to it.

Goals

Be a weekend project
Implement everything specific to ML/DL/NN from scratch
End up with a janky JS/WASM demo

Non-Goals

Be fast
Be good
Be clean code

Training

First, acquire the MNIST handwritten digit database (training/test sets, both images and labels) and save them in this repo.

Then, cargo run --release to run training, which will generate a net.pb containing the trained model.

Building web app

You'll need wasm-pack (cargo install wasm-pack). Then:

    cd bigbrainjs
    wasm-pack build --release --target web

You can then serve files from the bigbrainjs/web repository to see the web interface. The simplest way to do that is probably to run python -m http.server.

License

This work is free. You can redistribute it and/or modify it under the terms of the Do What The Fuck You Want To Public License, Version 2, as published by Sam Hocevar. See the COPYING file for more details.