Build a minimal checkpoint dict with the given class count. Args: num_classes: Total classes including background (bias shape). n

Write a minimal legacy .pth checkpoint and return its path.

Create a deterministic 100x100 binary mask with a rectangular region.

Write a minimal COCO annotation file. Args: path: Destination JSON path. categories: COCO category objects. annotations:

r""" Input: - memory: bs, \sum{hw}, d_model - memory_padding_mask: bs, \sum{hw} - spatial_shapes: nlevel, 2 Output:

Generate a full synthetic dataset in COCO format. Args: output_dir: Directory where the dataset will be saved. num_images: Total

Infer a keypoint schema from an Ultralytics YOLO pose YAML file. Args: data_file: Path to a YOLO ``data.yaml`` or ``data.yml`` file.

Return a [1, 4] float32 box tensor from (x1, y1, x2, y2).

Mock LightningModule whose ``device`` is a real string so ``metric.to(device)`` works.

Load a TFLite model, allocate tensors, and log I/O shapes. Tries ``tflite_runtime`` first (lightweight; preferred on edge devices), then falls ba

Return a _DummyRFDETR whose inner model carries custom class_names and returns given labels.

Build a minimal RFDETR-like namespace with controllable patch_size/num_windows.

Write a minimal COCO keypoint annotation file.

Minimal single-group AdamW — bypasses get_param_dict.

Convert RF-DETR keypoint precision parameters into pixel covariances. The keypoint head predicts lower-triangular precision-Cholesky parameters i

Return the training DataLoader. Uses a replacement sampler when the dataset is too small to fill ``_MIN_TRAIN_BATCHES`` effective batches

Call build_o365_raw with mocked CocoDetection and transform builders.

Detect the best available device **without** initialising the CUDA runtime. ``torch.cuda.is_available()`` creates a CUDA driver context that make

Resolve the list of transformer blocks/layers from backbone[0].encoder. Supports multiple backbone architectures: - encoder.blocks (s

Load images from a directory and prepare them for calibration. Images are loaded, resized to ``(height, width)``, converted to ``float32`` in ``[

Build a fake checkpoint dict.

Return a CocoEvaluator mock that returns the given stats array. Stats list must have exactly 10 elements matching _summarizeKps() output shape.

Create a minimal mock RFDETRModule with ``train_config.epochs``.

Create a minimal mock Trainer with controllable step metadata.

Fake legacy checkpoint with argparse.Namespace args.

Resize and ImageNet-normalise an image to match ``RFDETR.predict()``. Uses ``torchvision.transforms.functional`` when importable for bit-exact pa

Render the overall metrics table with merged group-header cells. Uses only plain Unicode box-drawing characters (no ANSI colour codes) so the out

Reset accumulated keypoint predictions for *split*. Args: split: One of ``"train"``, ``"val"``, ``"val_ema"``, or ``"test"``.

Resolve and validate the ``patch_size`` argument for :meth:`RFDETR.export` and :meth:`RFDETR.predict`. Args: patch_size: Value supplied b

Construct DataModule and inject a fake _dataset_train of given length.

Slice a ``refpoint_embed`` / ``query_feat`` weight preserving per-group structure. ``LWDETR`` packs query embeddings as ``nn.Embedding(num_querie

Return split-prefixed columns whose name contains any of the given patterns.

Create a minimal YOLO pose dataset on disk.

Build an LWDETR model directly from a ModelConfig. A config-native alternative to ``build_model(build_namespace(mc, tc))``. Constructs the namesp

Draw a geometric shape on an image and return its COCO polygon. The polygon is computed first, then used for both rendering and annotation, so th

Return lightweight metadata without loading pixels or dense masks.

Shared HungarianMatcher instance.

Build per-step rate arrays from trainer metadata. Args: trainer: The Lightning Trainer instance. pl_module: The ``RFD

Compute and log mAP and F1 metrics at the end of the validation epoch. Args: trainer: The PTL Trainer. pl_module: The

Unpack augmented boxes (and optionally masks), clamp to image bounds, remove zero-area boxes. After Kornia augmentation the padded ``[B, N_max, 4

Convert targets from normalised CxCyWH to absolute xyxy boxes. Also passes ``iscrowd`` and ``masks`` through unchanged. Args:

Upsample mask logits to image size and threshold at zero. Matches ``PostProcess.forward``: bilinear upsample with ``align_corners=False`` followe

Return a single-image target dict with one box in normalised CxCyWH.

Download a zip file and safely extract it into the destination directory. Args: url: URL to a zip archive. dest_dir: Directory wh

Put the underlying module in eval mode before unoptimized inference. Inference must never run with dropout / batch-norm in training mode. The

Load class names from a COCO or YOLO dataset directory.

Build a per-image prediction dict for MetricKeypointOKS.update(). Args: image_id: COCO image ID. kp_xy: List of (x, y) coordinate

Build a minimal param-dict list for AdamW from all trainable parameters.

Return a single-image prediction dict with zero-valued tensors.

Create a minimal RFDETR instance with mocked internals.

Return a minimal mock Trainer.

Return an overall dict with the minimal keys _render_overall_merged expects.

Prepare calibration data as a ``.npy`` file for ``onnx2tf``. The returned path points to a ``.npy`` file containing an NHWC float32 array with pi

Print two tables to the terminal: overall metrics and per-class metrics. The overall table is transposed (metrics as columns, one value row)

Serialize the model architecture config when the module exposes one. Schema-critical fields (``num_keypoints_per_class``, ``num_classes``) ar

Build a private diagnostic figure for transformed dataset samples. Samples the dataset after RF-DETR dataset transforms, so boxes and

Write a minimal valid COCO annotation file.

Compute the keypoint loss vector per matched target. The tensor layout follows GroupPose-style keypoints where each target class defines how

Write a Roboflow upload bundle (``weights.pt`` + ``class_names.txt``) into *output_dir*. This is the network-free core of :meth:`deploy_to_ro

Get ModelWeightAsset by filename. Args: filename: The model filename (e.g., 'rf-detr-base.pth') Returns: Mod

Pad a list of variable-size tensors into a single NestedTensor. Args: tensor_list: List of 3-D tensors (C, H, W) with possibly different

Factory fixture — call with **overrides to get a minimal SegmentationTrainConfig. tmp_path is injected automatically so test methods do not need

Strip a checkpoint file down to ``model``, ``args``, and PTL-compatible keys. Preserves ``model_name`` (when present) so that ``RFDETR.from_check

Aggregate keypoint class-logit contributions into detection-class logits.

Infer or validate keypoint schema from COCO, Roboflow COCO, or YOLO pose metadata. Args: config: Training configuration containin

Auto-detect the dataset class count and align ``model_config.num_classes`` in-place. Must be called before ``RFDETRModelModule`` is construct

Numpy bilinear resize matching ``F.interpolate(mode="bilinear", align_corners=False)``. Half-pixel center convention. Used by ``_decode_masks`` o

Build a single Albumentations transform from its name and parameter dict. Handles container transforms (``OneOf``, ``SomeOf``, ``Sequential``) by

Construct a DataModule with the given augmentation_backend.

Get a field from checkpoint ``"args"``, handling both dict and attribute access. New checkpoints (PTL training stack) store ``"args"`` as a plain

Compute MD5 hash of a file. Args: filepath: Path to the file. Returns: MD5 hash as hexadecimal string.

Download a file from a URL with optional MD5 validation. Args: url: URL to download from. filename: Path to save the file.

Encode a binary mask to compressed RLE with string counts (COCO JSON format).

Export a tiny model to ONNX and return the output file path. Args: tmp_path: Temporary directory provided by pytest. notes: Optio

Return a ``(C, H, W)`` float32 tensor filled with the given value.

Parse a config file by stem name and return its dict.

Return (1, 10, 4) array of normalised cxcywh boxes all centred at 0.5.

Return a simple TensorDataset of given length.

Build a minimal ``fit_loop`` state dict that restores the epoch counter. ``BestModelCallback`` stores ``trainer.current_epoch`` as ``"epoch"`` in

Validate that a file's MD5 hash matches the expected hash. Args: filepath: Path to the file. expected_md5: Expected MD5 hash.

Emit a ``logger.warning`` when ``load_state_dict`` left non-trivial gaps. ``load_state_dict(strict=False)`` silently ignores keys that the model

Build a Roboflow COCO-format dataset. This uses Roboflow's standard directory structure (train/valid/test folders with _annotations.coco.json).

For debug and test only, need to use cuda version instead.

Assert that randomized temporary download files were cleaned up.

Attach a mock trainer with the given training state to the DataModule.

Construct a DataModule for on_after_batch_transfer tests.

Compute pairwise boolean-mask IoU between N predictions and M ground truths. Args: pred_masks: Boolean mask tensor of shape [N, H, W].

Return a list with one per-image prediction dict.

Align out_features and out_indices against stage_names, filling in defaults when either is None. Copied from transformers.utils.backbone_utils.ge

Return True if *schema* uses a background-first layout. A background-first schema has a leading slot with zero keypoints that acts as a backg

Return a _DummyRFDETR whose model_config carries patch_size and num_windows.

Build a namespace for tests with minimal default TrainConfig values.

Build an RFDETR instance whose model/state are stubbed for export_for_roboflow. ``RFDETR.__init__`` is bypassed; only the attributes ``export_for