Dixon

An educational neural network library in Rust. It includes a simple MLP, dense layers, common activations, basic losses/metrics, and dataset loaders for Iris and MNIST. Examples demonstrate training, manual backprop, and model persistence.

Features

• MLP with per-sample SGD and MSE/CE losses
• Dense layers with configurable activations
• Numerically safe softmax + cross-entropy
• Iris CSV and MNIST (IDX gzip) loaders
• Save/load models to .pere (gzipped JSON) under models/
• Backprop API: compute gradients and apply updates

Workspace

• Library: dixon
• Examples: ml_examples

Quickstart

# Fetch datasets (MNIST to data/, Iris to data/)
bash scripts/download_datasets.sh

# Run only the Iris example
cargo run -p ml_examples --no-default-features --features iris

# Run only the MNIST example
cargo run -p ml_examples --no-default-features --features mnist

Library usage (basic)

use dixon::{MLP, Sigmoid};
use std::sync::Arc;

let mut mlp = MLP::new(4, vec![10, 5], 3, Arc::new(Sigmoid));
// Train with built-in loop or your own

// Save / load
mlp.save_pere("models/iris_model.pere")?;
let reloaded = MLP::load_pere("models/iris_model.pere")?;

Backprop API (custom training)

use dixon::Gradients;
let grads: Gradients = mlp.compute_gradients(&input, &target, "ce")?;
mlp.apply_gradients(&grads, 0.05);

Notes

• This is educational code: single-threaded, per-sample SGD. For best MNIST results, shuffle per-epoch (enabled), consider Linear output with CE (enabled in example), lower LR (e.g., 0.01), and run more epochs.
• MNIST loader looks in data/ first, then current directory.

Docs (GitHub Pages)

When pushed to master/main, API docs are published to GitHub Pages from gh-pages. Enable in repository Settings → Pages → Branch: gh-pages.

License

MIT