Suppress torch.jit.TracerWarning during export tests to reduce log spam.

Interpolate DINOv2 positional embeddings in *checkpoint_state* to match *pe_size*. When the model is configured with a custom ``resolution`` that

Return True if the current process is rank 0.

Entry point for the ``rfdetr`` CLI. Constructs and runs ``RFDETRCli`` with ``RFDETRModelModule`` and ``RFDETRDataModule``. Subcommands are auto-

Compute the bounding boxes around the provided masks. The masks should be in format [N, H, W] where N is the number of masks, (H, W) are the spat

Compute the distributed target-box denominator for a target batch. The denominator is the total number of ground-truth boxes in the batch, mu

Compute and log mAP and F1 under ``test/`` prefix at end of test epoch. Mirrors :meth:`on_validation_epoch_end` for the test evaluation loop.

Reset ``_ema_has_updates`` before test to prevent stale validation state from triggering EMA compute. ``on_test_batch_end`` never sets ``_ema

Optionally update EMA at epoch boundaries.

Compute optional train-split mAP at the end of the training epoch. Args: trainer: The PTL Trainer. pl_module: The Lig

Plot train/val/test detection and keypoint mAP metrics from a PTL ``metrics.csv`` file. Reads the CSV written by PyTorch Lightning's ``CSVLogger`

Resolve an ``augmentation_backend`` value to a concrete ``"cpu"`` or ``"gpu"``. ``"auto"`` resolves to ``"gpu"`` only when both CUDA and Kornia a

Switch the attention implementation without reloading the model. This is useful when you want to change the attention implementation after th

Print and log COCO summary statistics.

Sweep confidence thresholds and compute precision/recall/F1 at each. Args: per_class_data: Per-class matching data list indexed by class

Release the notebook output widget when the trainer exits. Args: trainer: The PTL Trainer. pl_module: The LightningMo

De-normalize a single image tensor, annotate it with boxes and labels, and plot it on ``ax``. Args: single_image: Normalized imag

Apply geometric transform to image with boxes and optionally masks. Converts data to Albumentations format, applies the transform, and conver

Fuse object scores with keypoint localization uncertainty. Args: scores_i: Object scores for one image before keypoint uncertaint

Populate missing COCO ``num_keypoints`` fields from visibility flags.

Convert a relative XYXY box into a 4-corner polygon.

Convert boxes to numpy array and validate shape. >>> import torch >>> boxes = torch.tensor([[10.0, 20.0, 30.0, 40.0]]) >>> Al

Flatten per-instance keypoints into Albumentations keypoint fields. >>> keypoints = np.array([[[10.0, 20.0, 2.0], [0.0, 0.0, 0.0]]], dtype=np

Build an in-memory ``pycocotools.COCO`` object from YOLO lazy samples. Args: classes: Ordered class names where index is the YOLO class I

Build a minimal COCO GT object with a single annotation. Args: num_keypoints: Number of keypoints per instance. gt_keypoints: Fla

Construct RFDETRDataModule (build_dataset is not called at init time).

Write a minimal ONNX model with one GridSample node to *path*.

Build category groups that can each be evaluated by one COCO keypoint evaluator.

Build a YOLO detection dataset that stores bounding boxes lazily. Unlike :func:`_build_lazy_yolo_segmentation_dataset`, this function does not st

Build a YOLO pose dataset that stores keypoints without dense masks.

Build a YOLO dataset that stores polygons and rasterizes masks on demand. Args: img_folder: Path to the directory containing images.

Build per-class rows and emit per-class AP metrics. Args: metrics: Output of ``MeanAveragePrecision.compute()``. pfx:

Return sorted COCO category ids that carry keypoint metadata or annotations.

Clear all per-instance fields when no boxes remain. >>> import torch >>> target = {"area": torch.tensor([100, 200]), "iscrowd": torch

Build a torchmetrics target dict from COCO ground-truth annotations for one image. Args: coco_gt: Loaded ``pycocotools.coco.COCO`` object

Return the native COCO train keypoint annotation path when it exists. Args: dataset_dir: Path to the COCO dataset root.

Module-level collate helper used as the base for :func:`make_collate_fn`. Defined at module scope (rather than as a closure inside :func:`make_co

Compute normalized losses for logging and raw weighted loss for backward. Args: outputs: Model output dictionary. tar

Return a mutable copy of a Supervision ``KeyPoints`` object.

Create attention mask that blocks cross-class keypoint interactions.

Load an ONNX model and create an ONNX Runtime inference session. Imports ``onnxruntime`` at call time so that the rest of the package remains usa

Create a minimal valid YOLO dataset directory structure.

Decode selected keypoints for one image into output tensors. Args: keypoints_i: Gathered selected keypoint predictions for one im

Detach postprocessed result tensors before handing them to callbacks. Args: results: Per-image postprocessed prediction dictionar

Detect the class count using the same category basis as training labels. For COCO-style datasets this counts all categories by ``id`` from ``

Convert batched supervision keypoints into the COCO evaluator format. Args: detections_batch: Per-image prediction batch returned by RF-D

Remove post-fit validation rows that Lightning logs as a synthetic final epoch. The Roboflow finetune demos run ``trainer.validate(...)`` after `

Shorten long sample titles so subplot grids do not overflow. Args: title: Raw title text, usually an image file name.

Create zero/NaN-filled keypoint output tensors for one image. Args: keypoints_i: Gathered selected keypoint predictions for one i

Encode a binary mask to uncompressed RLE with integer counts.

Return a faster-coco-eval COCO object for evaluator construction.

Run faster-coco-eval evaluation for accumulated COCO result records.

Evaluate one grouped keypoint result set.

Run keypoint evaluation per keypoint-count group and aggregate stats.

Extract and validate optional YOLO ``flip_idx`` metadata.

Extract YOLO keypoint names or synthesize stable placeholders.

Return one non-empty prediction per image so COCOEvalCallback has something to score.

Return a COCO object containing only the requested categories and annotations.

Filter a mapping-style augmentation config.

Filter per-instance fields to match kept box indices. >>> import torch >>> target = {"area": torch.tensor([100, 200, 300]), "iscrowd"

Convert a YOLO flip_idx permutation to flat swap pairs. Args: flip_idx: Full permutation where ``flip_idx[i]`` is the horizontal mirror o

Gather keypoint predictions for the selected query rows of one image. Args: out_keypoints_i: Keypoint predictions for one image w

Return an activation function given a string.

Return a built dataset split for private visualization.

Return the inner ``nn.Module`` wrapped by an EMA callback. ``RFDETREMACallback._average_model`` is a private attribute holding a ``torch.optim.sw

Resolve EMA model weights from the active EMA callback. Args: trainer: The Lightning Trainer instance. pl_module: The

Resolve live model weights from the active Lightning module. Args: pl_module: The ``RFDETRModelModule`` being trained. R

Return the patch-embedding projection layer for an RF-DETR model. RFDETR wrappers are not nn.Module; the underlying PyTorch module lives at ``mod

Infer a single keypoint count from COCO category metadata or annotations.

Insert a fake ``onnx2tf`` package into ``sys.modules``. Saves any pre-existing real modules under the same keys so they can be restored by ``_rem

Return whether a TrainConfig strategy string requests distributed execution.

Return True if transform (or any nested transform) affects spatial coordinates. For container transforms such as ``A.OneOf`` or ``A.Sequential``,

Return whether a metric column is a loss scalar.