Resize schema-dependent GroupPose state to match ``num_keypoints_per_class``.

Clone backbone projector weights into ``cross_attn_projector`` for dual-projector models. Older checkpoints can contain only ``backbone.0.project

Create and save image grids to ``output_dir``. Each grid is a 3x3 JPEG containing up to 9 images from a single batch, with bounding boxes and

Return True if *a* and *b* represent the same keypoint structure. Two schemas are semantically equal when they encode identical active keypoi

Emit *msg* as a WARNING exactly once per unique message string.

Return whether PTL auto/all device resolution can select multiple devices.

Accumulate a COCO evaluator and populate ``stats`` regardless of log mode.

Adapt a 3-channel pretrained conv weight tensor to *num_channels* input channels. When ``num_channels == 3``: returns the weight unchanged. When

Build a torchmetrics-compatible bounding-box dict from raw list or array data. Handles empty inputs transparently — an empty *boxes* list produce

Set up an :class:`~rfdetr.training.RFDETRDataModule` for validation. Calls ``setup("validate")`` so ``_dataset_val`` is ready. When *num_samples

Build module with a mock backbone that exposes backbone[0].encoder.

Build minimal valid inputs for ms_deform_attn_core_pytorch. Args: bsz: Batch size. n_heads: Number of attention heads. he

Copy pretrained weights from *rfdetr_obj* into a fresh :class:`~rfdetr.training.RFDETRModelModule`. Constructs the module with the same architect

Build a minimal :class:`~rfdetr.config.TrainConfig` for COCO inference runs. Loggers and EMA are disabled; the config is only used for validation

Calculate polygon area from COCO-style flat coordinates.

Compute pairwise BCE-with-logits summed along the last dim.

Compute a torchmetrics mAP metric while suppressing duplicate terminal summaries under progress bars.

Construct an RFDETRModelModule with all heavy work mocked. Returns the constructed module and the fake nn_model whose ``load_state_dict`` rec

Create an attention mask that blocks cross-class keypoint interactions.

Create a compact class-by-keypoint active mask for a keypoint schema.

Fake PTL-style checkpoint with dict args.

Extract contiguous YOLO class names from parsed YAML data.

Return a MagicMock that behaves enough like an LWDETR model.

Return *keys* with intentional-reinit/trim entries removed. Matching is boundary-aware: a pattern matches a key when the pattern appears at the s

Filter one augmentation entry, recursing into supported containers.

Return the EMA callback instance, or ``None`` if not present.

Return ``True`` when the runtime has a CUDA accelerator available. Uses the fork-safe global ``DEVICE`` constant from ``rfdetr.config`` so that t

Infer horizontal-flip swap pairs from left/right keypoint names.

Import class_path and construct an instance with init_args.

Build a COCO detections object, including the empty-result case.

Load a YAML file and require a mapping root. Args: yaml_path: Path to a YAML data file. Returns: Parsed YAML mapping. R

Log keypoint COCO stats from an already accumulated evaluator.

Return a namespace with all fields consumed by ``build_coco``.

Return a minimal checkpoint dict for _load_pretrain_weights_into tests.

Build a keypoint module wired with ``_ScalarLossModel`` and ``_BoxNormalizedCriterion`` for accum tests.

Build a bool _kp_active_mask tensor from a keypoints-per-class schema list.

Construct the smallest viable LWDETR without loading pretrained weights. Uses a MagicMock backbone and transformer with hidden_dim=4 so the model

Build a minimal PTL checkpoint with mismatched PE shape. Args: pe_size_src: Source grid side length (checkpoint was saved with th

Build an :class:`~rfdetr.training.RFDETRModelModule` from an RFDETR instance. Creates the module with the same architecture as *rfdetr_obj*, copi

Return the left/right counterpart for a normalized keypoint name.

Move model weights to the target device recorded in *model_ctx*. ``_build_model_context`` intentionally keeps the ``nn.Module`` on CPU so that ``

Parse one YOLO label line and return ``(class_id, xyxy_px, polygon_px)``. Args: values: Whitespace-split fields from the label line.

Place a compact multi-column legend below a matplotlib axes.

Plot columns with color by metric name and line style by split. When seaborn is installed, draws mean ± 1 std-dev bands computed from within-epoc

Run keypoint-preview inference in fixed-size batches. Args: model: Loaded keypoint-preview model. image_paths: COCO image paths t

Ensure ``map_metric_ema`` exists (and is reset) on EVERY rank when EMA is active. Driven by the rank-invariant presence of the EMA callback r

Resize and normalise a PIL image to an ``(1, C, H, W)`` float32 NCHW tensor. Resizes using ``BILINEAR`` to match ``torchvision.transforms.functio

Run one forward + loss + backward; return True if successful, False on OOM.

Raise a descriptive ValueError for a patch_size incompatibility. Args: ckpt_patch_size: patch_size recorded in (or inferred from) the che

Return the rfdetr console_scripts value from pyproject.toml.

Remove fake ``onnx2tf`` entries from ``sys.modules`` and restore originals. Args: saved: Snapshot returned by ``_install_fake_onnx2tf``.

Return whether Albumentations replay metadata applied a horizontal flip. Args: replay: ``ReplayCompose`` metadata from an Albumen

Skip test when current backend does not support grid_sample for dtype.

Reset the F1 accumulator for a metric split.

Reset keypoint precision-Cholesky output rows to unit Gaussian values.

Return a parameter with the first dimension resized by tiling or truncating rows.

Resolve checkpoint model_name from model_config or config type. The CLI/PTL path does not call ``RFDETR.train()``, so ``model_config.model_na

Resolve the Lightning precision string for a mocked device capability and ``amp_dtype``. Args: tmp_path: pytest temporary directo

Round *value* up to the next multiple of *multiple*. Args: value: Non-negative integer to round. multiple: Positive integer divis

Run ``RFDETR.predict()`` on a COCO val subset and return ``(mAP@50, macro-F1)``. Loads images from disk as PIL images, calls ``rfdetr_obj.predict

Load a deterministic subset of COCO person-keypoint validation images. Args: images_root: Directory containing COCO validation images.

Force the removed-module finder to behave as if shim files no longer exist.

Swap live model weights with averaged EMA weights.

Prepend a background slot to an active-first schema. A no-op when *schema* already starts with a zero-keypoint slot or is empty. Args:

Reference implementation: ``F.interpolate`` with ``align_corners=False``.

Warn once per keypoint count when using uniform custom OKS sigmas.

Write a minimal Roboflow COCO keypoint annotation file.

Convert boxes from [x1, y1, x2, y2] to [x1, y1, w, h].

Return the YOLO data file path when a dataset root has one. Args: dataset_dir: Path to the YOLO dataset root. Returns:

Compute pairwise IoU and union for two sets of boxes. Returns: iou: the NxM matrix containing the pairwise IoU values for eve

Calculate how much of a bounding box is outside the image boundaries. Args: bbox: Bounding box in [x_min, y_min, x_max, y_max] format.

Return ``(tensors, mask)`` tuple. Returns: Tuple of the padded tensor and the boolean mask.

Export mode.

Expose EMA model weights for external checkpoint callbacks.

Infer the keypoint schema stored in a checkpoint state dict.

Return the rank of the current process in the distributed group.

Return a short status string for the current git repo, or 'unknown' if unavailable. Returns: String describing the current git HEAD, stat

Get download URL for a model by filename. Args: filename: The model filename Returns: URL string if found, N

Get the current version of the specified package. Args: package_name: The name of the package to get the version for. Defaults to ``'rfde

Return merged cell width for columns start..end inclusive.