Add decoder-only deformable neck to Hungarian DETR detector

[Copilot]

The Hungarian DETR path was consuming backbone features too directly, without a deformable-transformer-style neck. This update introduces a decoder-only deformable neck (no encoder stack) and makes neck depth configurable for DINO backbone outputs.

• Decoder-only deformable neck

  • Added DeformableDecoderNeck and DeformableDecoderLayer in hungarian_head.py.
  • The neck builds multi-scale memory from backbone features, applies level embeddings + positional encoding, and refines learnable queries through a decoder stack only.

• Head integration

  • HungarianDetectionHead now routes backbone features through the new neck before class/box prediction.
  • Removed the prior direct TransformerDecoder path that operated on raw projected features.

• Configurability surfaced in detector API

  • DINOv2HungarianDetector now forwards:
    • num_decoder_layers (neck depth)
    • num_feature_levels (multi-scale memory levels)
  • replace_head(...) preserves these settings when rebuilding the head for class-count changes.

• Behavioral intent

  • Keeps the backbone as the encoder and uses decoder-only processing in the neck, aligning with deformable DETR-style query decoding for detector heads.

detector = DINOv2HungarianDetector(
    num_classes=10,
    model_name="facebook/dinov2-base",
    hidden_dim=256,
    num_queries=100,
    num_decoder_layers=8,   # configurable neck depth
    num_feature_levels=4,   # configurable multi-scale levels
    device="cuda",
)

[TNodeCode]

Can you also add a custom Detektion head that uses hungarian matching for matching detections and ground truth?

[Copilot]

Pull request overview

This PR integrates HuggingFace Transformers DINOv2 backbones (ViT and ConvNeXt) into the project and wires them into torchvision RetinaNet detectors, enabling training/inference with these newer feature extractors.

Changes:

• Added feature_extractors.py implementing DINOv2 ViT/ConvNeXt extractors plus RetinaNet-compatible backbone adapters (5-level pyramids).
• Added two new AbstractDetector implementations using those backbones with torchvision’s RetinaNet.
• Updated dependencies to include transformers.

Reviewed changes

Copilot reviewed 2 out of 3 changed files in this pull request and generated 2 comments.

File	Description
requirements.txt	Adds transformers dependency required by the new DINOv2/ConvNeXt backbones.
feature_extractors.py	Introduces DINOv2-based feature extractors and backbone adapters that emit RetinaNet-compatible feature pyramids.
detectors.py	Adds new RetinaNet detector classes that use the new DINOv2 ViT/ConvNeXt backbones.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[TNodeCode]

Then use output_hidden_states=True