https://github.com/rozukke/rust-gan

Small demo of PyTorch and CUDA run from a Rust wrapper with Nix deployment.
https://github.com/rozukke/rust-gan

Last synced: 8 months ago
JSON representation

Small demo of PyTorch and CUDA run from a Rust wrapper with Nix deployment.

• Host: GitHub
• URL: https://github.com/rozukke/rust-gan
• Owner: rozukke
• License: mit
• Created: 2024-07-26T04:57:46.000Z (about 1 year ago)
• Default Branch: main
• Last Pushed: 2024-08-13T12:30:19.000Z (about 1 year ago)
• Last Synced: 2025-01-10T04:27:59.505Z (9 months ago)
• Language: Python
• Size: 300 KB
• Stars: 0
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# rust-gan

[![GitHub license](https://img.shields.io/github/license/rozukke/rust-gan.svg)](https://github.com/rozukke/rust-gan/blob/main/LICENSE)

[![Creator rozukke](https://img.shields.io/badge/Creator-rozukke-f497af.svg)](https://github.com/rozukke)

[![Made with Rust](https://img.shields.io/badge/Made%20with-Rust-b7410e.svg)](https://www.rust-lang.org)

This is a repository meant to demonstrate running a PyTorch script from a Rust binary with versioning and deployment using Nix.

## Installation

Tested working on WSL with `Ubuntu-24.04 LTS`. Please ensure that you have the latest version Nvidia driver installed on your

system (tested with driver version `560.70`). Starting a shell with `nix develop` should print "CUDA found!" to the console.

## Usage

For development purposes, use `nix develop` to start a dev shell with all required packages. Place an appropriate model file

into the `model` directory. One such model can be downloaded [here](https://drive.google.com/file/d/1fCKufxu-a0vewCP1Y_7DP_JmrPxBXYCF/view?usp=sharing).

It is recommended to use this model as a specific architecture is required.

### Inference with shell

- The executable is in the format `rust-gan [OPTIONS] `, where the device is either `cpu` or `gpu`. If CUDA is not

accessible, the device will be overridden to use CPU with a warning.

- Provide a path to your image using `--input` and a path to save to with `--output`. Alternatively, run without arguments

to use the default input provided (found in `input/lowres.png`, output saved to `input/highres.png`).

- Provide a path to the model using a `--model` flag. Automatic model detection is not very compatible with nix packaging,

but might be added later.

- Use half precision with the flag `--half-precision`. Will be slower on CPU than full precision.

### Using Nix package

- It should be possible to run the nix package using `nix run ./ -- gpu --model /path/to/model`. This is currently tested with a local clone

(remote flake may not work as expected).

> [!IMPORTANT]  

> There seems to be a bug with some recent nightly Nix releases regarding local paths. To fix the package not building,

> the solution seems to be to force input fetching. `cannot fetch input 'path:./pysrc...`