Fix halo implementation and tiling artefact

[qin-yu]

Fix halo implementation and tiling artefact

1. Fix halo implementation from mirror padding to actual groundtruth
2. Fix normalisation from patch-level intensity normalisation to image-level
3. I recommend batch norm for training models for wide applicability

See my fix to pytorch-3dunet wolny/pytorch-3dunet#113 for more details.

[qin-yu]

I just fixed the problem in workflow:

• Conda build workflow fails because of branch name has / or - #219

But it seems that I need to add and modify tests

[lorenzocerrone]

Hey @qin very nice, I like the pydantic validation a lot.. should we also add it to the plantseg configs?
I think it would be so much nicer than the buggy validation I made.

!!!wrong pr!!! This comment was for the ModelZoo PR

[qin-yu]

Hey @lorenzocerrone, I

• fixed the wrong halo implementation,
• changed intensity normalisation during inference to be image-level instead of patch-level,
• fixed the misuse of branch name in conda build workfllow

Please have a look

[lorenzocerrone]

Looks good we can merge it :)

[qin-yu]

I improved the docstring based on Adrian's feedback: wolny/pytorch-3dunet#113 (comment)

Since I'm already improving the docstrings, I also fixed the arg types and made some simple improvement of code.

I'll merge this and work on the next.

[qin-yu]

Removing the unused argument in get_test_augmentations() introduced a bug. Then I found that I might also have overlooked the previous global normalisation. Going to check it later.