Any backbone for linear segmentation

[liopeer]

What has changed and why?

This is the beginning of a refactoring effort

• adds a new way for configs
• allows easy extension of linear segmentation for more backbones

Follow-ups will rename things away from DINOv2 and to generic names. After that I will move on to object detection, but don't want this to interfere with #772 etc. (ECViT implementations).

How has it been tested?

• I'll add unit tests later during clean-up (there are anyway renamings still missing etc.)

Did you update CHANGELOG.md?

• Yes
• Not needed (internal change)

Did you update the documentation?

• Yes
• Not needed (internal change without effects for user)

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[Copilot]

Pull request overview

This PR introduces a new “linear semantic segmentation” task-model implementation that is intended to be extendable across multiple backbone families by adding a small model-name → backbone/config registry, and it refactors internal model-wrapper typing to distinguish multi-scale ViT vs CNN capabilities.

Changes:

• Added a new linear semantic segmentation task model (LinearSemanticSegmentation) with tiling/untile inference utilities and a dedicated training loop.
• Introduced a registry-based configuration layer for mapping *-linear model names to a backbone + default backbone args.
• Refined model wrapper / package typing to express “multi-scale feature” capability for ViTs vs CNNs, and updated DINOv2/DINOv3 wrappers accordingly.

Reviewed changes

Copilot reviewed 15 out of 15 changed files in this pull request and generated 21 comments.

Show a summary per file

File	Description
src/lightly_train/_task_models/linear_semantic_segmentation/transforms.py	Adds default train/val transform arg presets for linear semantic segmentation.
src/lightly_train/_task_models/linear_semantic_segmentation/train_model.py	Adds the training/validation loop for linear semantic segmentation.
src/lightly_train/_task_models/linear_semantic_segmentation/task_model.py	Adds the inference-time task model (tiling, postprocess) and backbone wiring.
src/lightly_train/_task_models/linear_semantic_segmentation/config.py	Adds registry entries mapping *-linear model names to backbone definitions.
src/lightly_train/_models/package.py	Introduces MultiScaleFeaturePackage and a model-name parsing protocol.
src/lightly_train/_models/model_wrapper.py	Splits multi-scale feature protocols into ViT/CNN variants and adds stride/patch-size capability.
src/lightly_train/_models/dinov3/dinov3_vit.py	Updates DINOv3 ViT wrapper to provide patch size and new protocol typing.
src/lightly_train/_models/dinov3/dinov3_convnext.py	Updates DINOv3 ConvNeXt wrapper typing and adds multi-scale stride reporting.
src/lightly_train/_models/dinov3/dinov3_src/models/convnext.py	Adds a type annotation for downsample_layers.
src/lightly_train/_models/dinov3/dinov3_package.py	Marks DINOv3 as a multi-scale-feature package and adds protocol typing.
src/lightly_train/_models/dinov2_vit/dinov2_vit.py	Updates DINOv2 ViT wrapper to provide patch size and new protocol typing.
src/lightly_train/_models/dinov2_vit/dinov2_vit_package.py	Marks DINOv2 as a multi-scale-feature package and adds protocol typing.
src/lightly_train/_configs/model_registry.py	Adds a generic registry utility used for model/config lookup.
src/lightly_train/_configs/config.py	Adds ConfigsNamespace to enforce “namespace-only” config containers.
src/lightly_train/_commands/train_task_helpers.py	Switches the training task helper to import the new linear semantic segmentation train model.

[liopeer]

Note: Eventually all packages should implement this.

[liopeer]

I'll do a lot of renaming from dinov2_linear_semantic_... to just semantic_ in a follow up.

[liopeer]

/review