feat: incorporate ECViT backbones into DINOv3 LTDETR

[gabrielfruet]

Summary

Incorporates the EdgeCrafter ECViT backbones into the DINOv3 LTDETR object detection task model, exposing four new model names under the edgecrafter/ package prefix:

• edgecrafter/ecvitt-ltdetr
• edgecrafter/ecvittplus-ltdetr
• edgecrafter/ecvits-ltdetr
• edgecrafter/ecvitsplus-ltdetr

Each is a drop-in replacement for the DINOv3 ViT LTDETR backbones — only the model argument changes:

lightly_train.train_object_detection(
    out="out/my_experiment",
    model="edgecrafter/ecvitt-ltdetr",  # only change vs. dinov3/*-ltdetr
    data={...},
)

Stacking

This PR stacks on #772 (gabriel-trn-2143-port-ecvit-to-lightlytrain), which moved the ECViT backbone port into src/lightly_train/_models/ecvit/. #772 must be merged first. The base branch for this PR is gabriel-trn-2143-port-ecvit-to-lightlytrain, not main.

Design

• Incorporation strategy: extended the existing DINOv3LTDETRObjectDetection task model to dispatch on package_helpers.parse_model_name(model_name). Accepts both "dinov3" and "edgecrafter" packages. No new task model class.
• No new encoder/decoder configs. The ECViT wrapper emits (P3, P4, P5) at strides 8/16/32 with per-level channel counts equal to proj_dim (192 for ecvitt, 256 for the other three). These match the existing DINOv3 ViT-shaped HybridEncoder / RTDETRTransformerv2 / DFINETransformer configs by proj_dim, so the configs are reused:
  • ecvitt → _DINOv3LTDETRObjectDetectionViTTConfig (in_channels 192)
  • ecvittplus / ecvits / ecvitsplus → _DINOv3LTDETRObjectDetectionViTTPlusConfig (in_channels 256)
• use_sta=False. ECViT already fuses its own pyramid internally; no SpatialPriorModulev2 is instantiated. The wrapper is a pass-through: forward(x) → ECViTWrapper(x).
• No mask_token freeze. ECViT has no mask_token; the DINOv3 ViT branch's backbone.mask_token.requires_grad = False line is skipped.
• Fixed patch_size=16. ECViT-NN uses a ConvPyramidPatchEmbed that only supports patch_size=16. The train/val transform's image size and scale_jitter.divisible_by are resolved to match (multiples of 32).
• Fixed _expected_input_channels=3. Multi-channel input is intentionally not supported for ECViT on this first cut. The constructor ignores backbone_args["in_chans"] and image_normalize-driven channel counts on the edgecrafter branch.
• Pretrained weights. Loaded via ECVIT_PRETRAINED_URLS through the new EdgeCrafterPackage.get_model, which downloads to Env.LIGHTLY_TRAIN_MODEL_CACHE_DIR and loads into ECViTWrapper.backbone via the existing _load_backbone_weights path. No hash check on first cut (matches the DINOv3 TODO).

Files

New (3):

• src/lightly_train/_models/ecvit/ecvit_package.py — EdgeCrafterPackage (public name "edgecrafter", alias ECVIT_PACKAGE)
• src/lightly_train/_task_models/dinov3_ltdetr_object_detection/ecvit_vit_wrapper.py — ECViTBackboneWrapper (pass-through adapter)
• tests/_task_models/dinov3_ltdetr_object_detection/test_ecvit_vit_wrapper.py — parametrized forward-shape parity, patch_size=16, backbone_model identity, no mask_token

Edited (4):

• src/lightly_train/_models/package_helpers.py — register EDGE_CRAFTER_PACKAGE
• src/lightly_train/_task_models/dinov3_ltdetr_object_detection/task_model.py — dispatch, list_model_names, parse_model_name, constructor
• src/lightly_train/_task_models/dinov3_ltdetr_object_detection/train_model.py — resolve_auto ECViT branch → patch_size=16
• tests/_task_models/dinov3_ltdetr_object_detection/test_task_model.py — ECViT parametrizations (is_supported_model, parse_model_name, _create_train_model, set_train_mode, resolve_auto)

No Makefile change required: the blanket add-header walk stamps the Lightly header on every .py under src/ unless excluded by -x, and the new files are not excluded.

Verification (local, Python 3.13)

• make format ✓
• make static-checks ✓ — mypy: "Success: no issues found in 557 source files"
• Focused tests: 63 passed, 2 skipped in 9.67s (uv run --frozen pytest tests/_task_models/dinov3_ltdetr_object_detection/test_ecvit_vit_wrapper.py tests/_task_models/dinov3_ltdetr_object_detection/test_task_model.py -q). The 2 skips are pre-existing ONNX tests that need onnx/onnxruntime/GPU.
• make test was intentionally skipped locally per the established pattern (CI runs it).

Out of scope (intentional)

• No documentation changes in this PR. The edgecrafter/ package page (docs/source/pretrain_distill/models/edgecrafter.md), the toctree/bullet entry in docs/source/pretrain_distill/models/index.md, and the "LTDETR EdgeCrafter (ECViT) Models" subsection in docs/source/object_detection.md are deferred to a separate docs-only follow-up PR stacked on top of this one.
• No ECViTWrapper → ECViTModelWrapper rename in this PR. The backbone-level naming standardization will land in a separate follow-up PR.
• No *-ltdetr-coco rows in the docs table — we have no COCO-fine-tuned ECViT weights yet.
• No mAP numbers in the docs table.
• No Quick Start snippet for ECViT in object_detection.md (the DINOv3 snippet applies verbatim; only the model argument changes).
• No multi-channel support for ECViT (always 3 input channels).
• No new _HybridEncoder* / _RTDETRTransformerv2* / _DFINETransformer* configs — the existing DINOv3 ViT-shaped configs are reused.
• No new ECViT-specific backbone-wrapper config (_RTDETRBackboneWrapperViTT*Config) — those are STA-specific and we have use_sta=False.

Related

• Parent PR: feat: add ECViT LTDETR backbone wrapper #772 (gabriel-trn-2143-port-ecvit-to-lightlytrain) — moves the ECViT backbone port into src/lightly_train/_models/ecvit/. Must be merged first.
• Upstream: EdgeCrafter (capsule2077/edgecrafter).
• Task model: DINOv3LTDETRObjectDetection (the class name is unchanged; ECViT is dispatched internally).

Test plan for reviewer

1. uv run --frozen pytest tests/_task_models/dinov3_ltdetr_object_detection/ -q → expect 63 passed, 2 skipped.
2. uv run --frozen pytest tests/_models/ecvit/ -q → expect existing ECViT tests to still pass (no changes to ecvit.py).
3. lightly_train.list_models() should include the four new edgecrafter/ecvit*-ltdetr names.
4. lightly_train.load_model("edgecrafter/ecvitt-ltdetr") should download pretrained weights and instantiate without error (requires network for the first run).