Reuse of metrics and losses

[berombau]

This library's metrics and losses are great for instance and panoptic segmentation! Have you considered creating a separate library for them or making reuse in other projects easier? For example, making bigger or limiting dependencies like numba, opencv-python, geopandas, pathos... optional.

The only dependencies for the metrics seem to be numpy and for the losses torch, but maybe I'm missing some utils functions.

Problems I have with other libraries:

• TorchMetrics
  • focusses more on optimization losses. Could be a place to upstream some of the losses and improvements in this project, they already have some detection and semantic segmentation losses currently.
• MONAI
  • Extensive and well supported implementation for torch losses and metrics.
  • Also suports MetricsReloaded, with a nice and very complete paper for metrics, but the code base and API is not that straightforward or lightweight to work with e.g. this seems to be the main tutorial.

[okunator]

Good suggestion, thanks! I can for sure create a standalone library for the metrics.