- GPU/CPU selection not respected
- Error messages not helpful in many cases
- Filter-out broken models?
Update
BioImageIO does not handle dependencies automagically, so many models do not work in panseg atm. Models do specify their dependencies and there is the get_conda_env function in bioimageio.spec, so we could implement supporting it, with the risk of cluttering/breaking the panseg environment.
Device selection is supported only for custom prediction pipelines using core._prediction_pipeline.create_prediction_pipeline, which we currently do not use.
Still, memory would not be freed in some cases, making it suboptimal for trying out multiple models as users might want to, given a nice UI.
Update
BioImageIO does not handle dependencies automagically, so many models do not work in panseg atm. Models do specify their dependencies and there is the
get_conda_envfunction inbioimageio.spec, so we could implement supporting it, with the risk of cluttering/breaking the panseg environment.Device selection is supported only for custom prediction pipelines using
core._prediction_pipeline.create_prediction_pipeline, which we currently do not use.Still, memory would not be freed in some cases, making it suboptimal for trying out multiple models as users might want to, given a nice UI.