Add segger (Draft)

[LouisK92]

This is a first draft for adding Segger. There are current issues that can't be solved (spatialdata and segger dependency issue and cuda error when running on cpu)

Two main issues:

• When running the last step (segment) we run into a cuda issue (tested in dockercontainer on Mac M3), note that the training worked fine. Related issue: CUDA error when running predictions without GPU EliHei2/segger_dev#126
• The latest segger version is needed for (hopefully a future solution of) the cuda issue, and the support of other techs than xenium (no need for the "qv" column, see here Data sample save does not produce files EliHei2/segger_dev#125). However, there are dask und numpy version dependency conflicts between segger and spatialdata (see here Spatialdata and dask Comatibility EliHei2/segger_dev#123).

Environment

• For the docker container, note that the config.vsh.yaml gives the ideal solution (for future updated version of the danielunyi42/segger_dev container), however currently this doesn't work (see CUDA error when running predictions without GPU EliHei2/segger_dev#126 (comment)). Besides the docker container build that I posted there you also need to install:

git clone https://github.com/EliHei2/segger_dev.git
cd segger_dev
pip install .
pip install torch==2.5.0 
pip install lightning
pip install torchmetrics
pip install cupy-cuda12x
pip install distributed
pip install dask_geopandas

Other

• There are a few NOTEs and TODOs in the segger script that can be tackled when the upper issues are resolved

How to proceed

• Continue when the spatialdata segger dependency issue is resolved
• Hope that the cuda error is then also resolved

[LouisK92]

Update: the spatialdata dask dependency issue is probably fixed with spatialdata>=0.6
Currently we still use <0.6 since sopa does not support zarr v3 yet. When sopa supports v3, we can update everything to the latest spatialdata version. Note that the datasets need to be recreated then (from the combine script onwards)