Add EUPE support

[guarin]

What has changed and why?

Adds support for the new EUPE models from Meta.

Usage is very simple. Instead of model="dinov3/vits16" for pretraining or model="dinov3/vits16-ltdetr" for fine-tuning you can now use model="dinov3/vits16-eupe" for pretraining with EUPE or model="dinov3/vits16-eupe-ltdetr" for fine-tuning.

I'll add docs/changelog in a follow up once #691 is merged.

Distillation

import lightly_train

if __name__ == "__main__":
    lightly_train.pretrain(
        out="out/my_experiment",
        data="my_data_dir",
        model="dinov3/vitt16",
        method="distillation",
        method_args={
            "teacher": "dinov3/vitb16-eupe", # Distill from EUPE ViT-B/16
        },
    )

Fine-tuning

import lightly_train

if __name__ == "__main__":
    lightly_train.train_semantic_segmentation(
        out="out/my_experiment",
        model="dinov3/vits16-eupe-eomt",
        data={...},
    )

How has it been tested?

• Manually

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 Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e049c9c590

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[Copilot]

Pull request overview

Adds support for Meta’s EUPE checkpoints in the DINOv3 model package and ensures EUPE backbones can be used with the DINOv3 LTDETR object detection task.

Changes:

• Register new *-eupe DINOv3 backbone names and point them at EUPE checkpoint URLs.
• Make LTDETR config selection tolerant to -eupe suffixes on backbone names.
• Strip EUPE-specific projector weights from loaded checkpoints to keep strict state-dict loading working.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 5 comments.

File	Description
src/lightly_train/_task_models/dinov3_ltdetr_object_detection/task_model.py	Strips -eupe from backbone names when selecting the LTDETR config.
src/lightly_train/_models/dinov3/dinov3_src/hub/backbones.py	Filters EUPE-only state-dict keys before strict loading for ViT and ConvNeXt backbones.
src/lightly_train/_models/dinov3/dinov3_package.py	Adds EUPE model-name entries mapping to existing backbone builders + EUPE checkpoint URLs.

[guarin]

/review

[liopeer]

LGTM! I will create a follow-up to add support for other downstream tasks and to document it.

[guarin]

LGTM! I will create a follow-up to add support for other downstream tasks and to document it.

All downstream tasks should already be supported. LTDETR just needed some extra logic