Test case for the tract library
This is an experiment on how to wrap a PyTorch-trained ONNX U-Net model into a standalone Rust executable.
- Input/Output image size is fixed for now (1211 x 900), but arbitrary sizes can be extended easily.
- Execution is entirely single core! That means rather high latency, but almost linear scalability.
- The model could be embedded in the executable as well using the include_bytes! macro.
- For the curious: The bundled mini model does some (bad) pixel segmentation on binary sheet music images. The full model is alot bigger and better.