Neural Numbers (Rust + eframe/egui + tch-rs)

A small desktop app that recognizes handwritten digits (0–9).

This project is primarily a practical playground to learn:

• building a GUI in Rust with eframe/egui
• converting mouse strokes into a MNIST-like 28×28 tensor
• training / saving / loading weights with tch-rs (libtorch)
• fine-tuning on your own samples collected from the GUI

---

Binaries

• solver_app — GUI: draw a digit, get prediction + probabilities, save correct samples
• trainer_app — CLI: train on MNIST and fine-tune on mydata/

---

Screenshots

---

Features

Solver (GUI)

• Mouse drawing canvas
• Brush size slider
• Space hotkey → clear canvas
• Live prediction + per-class probabilities
• Save your correct label to build a dataset: mydata/<0..9>/*.bin

Trainer (CLI)

• Loads MNIST from data/ (IDX ubyte format)
• Trains a small MLP (784 → hidden → 10)
• Saves weights to models/mnist.ot (configurable)
• Fine-tunes on your GUI samples from mydata/

---

Quick start

Run solver (GUI)

cargo run --bin solver_app

Run trainer (MNIST training + fine-tune)

cargo run --bin trainer_app

Tip: both binaries read config.json from the current working directory. Run from the repo root to use the default paths (data/, models/, mydata/).

---

Config (config.json)

Both binaries read/create config.json in the current working directory.

Example (default in this repo):

{
  "epoch": 300,
  "data_dir": "data/",
  "out_path": "models/mnist.ot",
  "image_dim": 784,
  "hidden": 128,
  "labels": 10
}

Notes:

• image_dim is 28×28 = 784 (MNIST-like flattened input).
• hidden is the MLP hidden layer size.
• out_path is what Solver loads and what Trainer writes.

---

Data formats

MNIST (data/)

Trainer expects the classic 4 MNIST IDX files in data/:

• train-images-idx3-ubyte
• train-labels-idx1-ubyte
• t10k-images-idx3-ubyte
• t10k-labels-idx1-ubyte

My samples (mydata/)

Solver saves your drawings as raw 28×28 grayscale:

• Path: mydata/<label>/<timestamp>.bin
• Each .bin is exactly 784 bytes (u8), row-major, values 0..255

Trainer reads this folder and fine-tunes the model.

---

Project structure

• src/bin/solver_app.rs — GUI entry point (eframe::run_native)
• src/bin/trainer_app.rs — training entry point

Shared

• src/bin/shared_lib/c_trainer_config.rs — config load/save (config.json)
• src/bin/shared_lib/f_ai_data.rs — model builder (MLP)

Solver

• src/bin/solver_lib/c_solver_app.rs — UI + hotkeys + “save sample”
• src/bin/solver_lib/c_painter_module.rs — stroke collection + canvas
• src/bin/solver_lib/f_utils.rs — rasterize → 28×28 + predict + save helpers

---

Notes / troubleshooting

GPU / CUDA

The code uses:

Device::cuda_if_available()

If you still see CPU, your libtorch build is likely CPU-only. To use CUDA, you typically need a CUDA-enabled libtorch and/or to point tch-rs to an existing PyTorch install.

---

About

A minimal “draw → rasterize → normalize to MNIST 28×28 → infer → fine-tune” pipeline in pure Rust.