refactor: extract load_pretrain_weights + apply_lora into models/weights.py

[Borda]

What does this PR do?

Eliminates duplicated weight-loading and LoRA application logic between detr.py (_load_pretrain_weights_into, _apply_lora_to) and module_model.py (_load_pretrain_weights, _apply_lora). The canonical unified implementation lives in the new src/rfdetr/models/weights.py and both callers delegate to it.

Key changes:

• New src/rfdetr/models/weights.py: load_pretrain_weights() uses the more complete module_model.py logic (Pydantic model_fields_set for user-override detection, auto-align to fine-tuned checkpoints, expand-back for explicit num_classes) plus class_names extraction from detr.py. apply_lora() is the unified LoRA application function.
• detr.py: removes ~80 lines of private weight/LoRA methods; delegates to weights.py; fixes mutable COCO_CLASS_NAMES returned directly from class_names property (now returns list copy).
• module_model.py: removes ~95 lines of private weight/LoRA methods; delegates to weights.py.
• tests/models/test_weights.py: 8 new characterization tests (4 reinit scenarios, 2 class_names extraction, 2 LoRA application).
• test_load_pretrain_weights.py, test_module_model.py, test_detr_shim.py: migrated patch targets and call sites to rfdetr.models.weights.

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• Refactoring (no functional changes)

Testing

• I have tested this change locally
• I have added/updated tests for this change

Additional Context

[Copilot]

Pull request overview

This PR centralizes pretrained checkpoint loading and LoRA application into a single shared module (rfdetr.models.weights) and updates both the inference facade (rfdetr.detr) and the training LightningModule (rfdetr.training.module_model) plus tests to use that unified implementation.

Changes:

• Added src/rfdetr/models/weights.py providing canonical load_pretrain_weights() and apply_lora().
• Refactored src/rfdetr/detr.py and src/rfdetr/training/module_model.py to delegate weight loading / LoRA application to rfdetr.models.weights.
• Updated and added tests to patch and validate the new unified code paths (including new characterization tests).

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
src/rfdetr/models/weights.py	New unified implementation for loading pretrained weights and applying LoRA.
src/rfdetr/training/module_model.py	Removes local helpers and calls the shared weights/LoRA utilities.
src/rfdetr/detr.py	Removes duplicate helpers, delegates to shared utilities, and returns a copy for COCO class-name fallback.
src/rfdetr/models/init.py	Re-exports load_pretrain_weights / apply_lora from rfdetr.models.
tests/models/test_weights.py	New unit tests directly exercising rfdetr.models.weights.
tests/training/test_module_model.py	Migrates patch targets and call sites to rfdetr.models.weights.
tests/training/test_load_pretrain_weights.py	Refactors regression coverage to exercise the unified loader directly.
tests/training/test_detr_shim.py	Migrates detr-path checkpoint compatibility tests to the unified loader and updates class_names identity assertions.

[Copilot]

pretrain_weights = mc.pretrain_weights can be None (ModelConfig allows it), but this function immediately passes it into download_pretrain_weights() / os.path.isfile(), which will raise a type error if a caller invokes load_pretrain_weights() without a configured path. Since this helper is now a public import, add an explicit guard (e.g., early return [] or raise a clear ValueError) when pretrain_weights is None.

[Copilot]

class_names = _ckpt_args_get(checkpoint["args"], "class_names") or [] doesn't normalize the result to a list[str]. _ckpt_args_get can return any type stored in the checkpoint (e.g., tuple, numpy array, or other sequence), which would violate the annotated return type and can leak a mutable reference if it's already a list. Consider coercing to a new list (and optionally filtering to strings) before returning.

Suggested change

	class_names = _ckpt_args_get(checkpoint["args"], "class_names") or []
	raw_class_names = _ckpt_args_get(checkpoint["args"], "class_names")
	if raw_class_names:
	# Normalize to a new List[str] to avoid leaking mutable references and
	# to respect the annotated return type.
	if isinstance(raw_class_names, str):
	class_names = [raw_class_names]
	else:
	try:
	iterator = iter(raw_class_names)
	except TypeError:
	# Non-iterable, ignore and keep the default empty list.
	class_names = []
	else:
	class_names = [name for name in iterator if isinstance(name, str)]

[Copilot]

When auto-aligning to a fine-tuned checkpoint (checkpoint_num_classes < configured_num_classes_plus_bg and the user did not override the default), the code updates only the local num_classes variable but does not persist the aligned value back to model_config.num_classes. That leaves model_config (and any pre-built namespaces based on it) inconsistent with the loaded model head for callers that don't separately sync from nn_model.num_classes (unlike RFDETRModelModule). Consider updating model_config.num_classes as part of the auto-align path (or returning the effective class count) so all callers stay consistent.

Suggested change

	configured_num_classes_plus_bg = checkpoint_num_classes
	configured_num_classes_plus_bg = checkpoint_num_classes
	# Keep the ModelConfig in sync with the effective class count.
	mc.num_classes = num_classes

[Copilot]

The docstring claims this function raises RuntimeError when the checkpoint cannot be loaded after re-download, but the implementation will re-raise whatever exception torch.load() throws on the second attempt. Either wrap the second failure and raise RuntimeError as documented, or adjust the docstring to match actual behavior.

Suggested change

	RuntimeError: If the checkpoint file cannot be loaded after a re-download.
	Exception: If the checkpoint file cannot be loaded even after a re-download.

[codecov]

Codecov Report

❌ Patch coverage is 94.66667% with 4 lines in your changes missing coverage. Please review.
✅ Project coverage is 75%. Comparing base (3cd6898) to head (dc2357b).

Additional details and impacted files

@@          Coverage Diff           @@
##           develop   #840   +/-   ##
======================================
  Coverage       75%    75%           
======================================
  Files           92     93    +1     
  Lines         7148   7134   -14     
======================================
- Hits          5383   5377    -6     
+ Misses        1765   1757    -8     

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.