Grayscale / 1ch input on LibreYOLO9

[wizmo2]

Asking to see if there is an existing way to implement grayscale models with LibreYolo9. The ''channel' or 'ch' yaml parameter seems to be ignored. Ultimately the need is for a custom trained 'onnx' single channel input model based on Yolo (9t,or26n) to run in .NET.

[EHxuban11]

Hello @wizmo2,

Thanks for asking! LibreYOLO didn't support this yet, so we added it in PR #1015: add channels: 1 to your data YAML (same key as Ultralytics) and train as usual. The exported ONNX then takes a [batch, 1, H, W] gray input.

It's on dev and will ship in the next release (v1.8.0) . Let us know if it works for you :)

[wizmo2]

Thank you for you quick response.

Tested on GPU (5060 / CUDA 13.2 / py 3.12.14) with LibreYOLO9t. Onnx model works as expected.

Great work!

[EHxuban11]

Excelent, happy to hear that ! I will close this discussion but don't hesitate to open other issues in the future if you encounter problems with grayscale or any other aspect of the library.