Check whether a file exists and contains at least one byte. Returns ``False`` for a missing or empty file so callers can trigger a re-downloa

Temporarily patch :func:`numpy.load` to set ``allow_pickle=True``. ``onnx2tf`` 1.x calls ``np.load()`` on its bundled calibration data without pa

Read a PTL CSVLogger metrics file and return both step-level and epoch-averaged DataFrames. Returns: A ``(raw_df, epoch_df)`` pair where

Rewrite every GridSample node in *onnx_path* to use TFLite-safe ops. ``onnx2tf``'s built-in GridSample handler lowers to ``tf.gather_nd(batch_dim

Rescale gradients already accumulated in the current optimizer window. Args: scale: Multiplicative factor that converts previous

Resolve ``"auto"`` to ``"cpu"`` or ``"gpu"`` based on runtime availability. For ``"cpu"`` and ``"gpu"`` the value is returned unchanged. For ``"

Run deploy.export.main(args) with all heavy dependencies mocked. Stubs out build_model, make_infer_image, and export_onnx, and injects mock o

Replace non-finite cost entries with a large finite sentinel. >>> HungarianMatcher._sanitize_cost_matrix( ... torch.tensor([[1.0,

Write a small solid-colour RGB JPEG to *path*.

Helper: construct DataModule and call setup(stage) with build_dataset mocked.

Suppress download/validate/file-existence side effects on the canonical load path.

Strip the leading background slot from a bg-first schema. Always returns a new list. A no-op (copy) when *schema* is already active-first or

Write a minimal COCO annotation file with provided categories.

Create a minimal Roboflow YOLO dataset root.

Export a model to ONNX. Args: output_dir: Directory where the ONNX file will be written. model: Model to export. input_na

Args: norm: Either one of BN, SyncBN, FrozenBN, GN; or a callable that takes a channel number and returns the normalization l

Return the number of processes in the current distributed group.

Checks if the specified dataset directory is in yolo format. We accept a dataset to be in yolo format if the following conditions are met: -

Skip weight download in tests.

Accumulate train predictions for optional train-split mAP logging. Args: trainer: The PTL Trainer. pl_module: The Lig

Apply the scheduled rates for the current global step. Args: trainer: The Lightning Trainer instance. pl_module: The

Read a PTL ``CSVLogger`` metrics file and build a training plot. The figure contains one subplot per metric group (loss, detection metrics, k

Run inference on a preprocessed batch and return postprocessed results. Args: batch: Tuple of (NestedTensor samples, list of targ

Resize the detection classification head to *num_classes* outputs. Replaces ``self.class_embed`` (and each ``enc_out_class_embed`` when the m

Clear accumulated predictions. Examples: >>> from unittest.mock import MagicMock >>> metric = MetricKeypointOKS(Magic

Return the test DataLoader. Returns: DataLoader for the test dataset with sequential sampling.

Move a ``(NestedTensor, targets)`` batch to *device*. PTL's default iterates tuple elements and calls ``.to(device)``; that works for plain t

Compute the EMA update for a single parameter tensor. Matches the ``ModelEma`` formula where ``updates`` is 1-indexed: PTL's ``num_averaged``

Build minimal detection outputs used across matcher keypoint tests.

Return a [1, h, w] boolean mask with the specified region set to True.

Return an [N, 4] float32 tensor from a sequence of [x1,y1,x2,y2] rows.

Map COCO category ids onto model label slots that have keypoint capacity. RF-DETR keypoint schemas are indexed by model label. The preview person

Build the class list and sample list shared by both YOLO builder functions. Iterates over every image in ``img_folder``, reads image dimensions v

Fake checkpoint with model_name key (new format).

Instantiate a minimal parser and collect registered link sources.

Compute and log OKS keypoint AP/AR metrics when keypoint mode is active. Args: split: Internal metric split (``"val"``, ``"val_em

Return a mock that ``_get_ema_callback`` recognises (has ``get_ema_model_state_dict``).

Read class names from a YOLO ``data.yaml`` file.

Return a MagicMock criterion with a realistic weight_dict.

Return a callable MagicMock for postprocess.

Filter ordered single-key augmentation entries.

Run 1 epoch (2 train + 2 val batches) and return the resulting metrics.csv.

Return ``True`` when the runtime has a CUDA accelerator available. Uses the fork-safe global ``DEVICE`` constant instead of direct ``torch.cuda.i

Infer keypoint count for each category from COCO category metadata or annotations.

Build a zero keypoint prediction tensor with the shape expected by _aggregate_keypoint_class_logits. The second-to-last dimension must equal num_

Return minimal args consumed by ``build_roboflow_from_coco`` in keypoint mode.

Build a checkpoint with labelled query weights and dict-style args.

Construct a minimal keypoint-capable LWDETR with detection head resized to num_classes+1. Mirrors what happens after loading a pretrained checkpo

Write a minimal checkpoint to a temp file.

Build minimal valid inputs for MSDeformAttn.forward. Returns: Tuple of (query, reference_points, input_flatten,

Build a module with a stub trainer exposing ``num_training_batches`` for the test scenario.

Build target dicts with boolean mask tensors for given instance counts.

Greedy highest-score-first matching for one class in one image. Implements the COCO matching algorithm: each GT is matched at most once; detectio

Return one global flip-pair list when all keypoint categories agree.

Mock MeanAveragePrecision carrying minimal per-image state lists (one entry each).

Plot mAP metrics on a single axes with line style by split.

Build a figure for grouped metrics.

Mirror of the PyTorch predict() preprocessing: to_tensor -> resize -> normalize.

Verify that Kornia is importable, raising a clear error if not. Raises: ImportError: When ``kornia`` is not installed, with an install hi

Resolve a predicted label to a COCO category_id.

Resolve ``augmentation_backend`` at runtime for dataset builders. Thin wrapper around :func:`rfdetr.datasets.kornia_transforms.resolve_augmentati

Clamp a MultiheadAttention head count to at least one.

Update class names and keep args.num_classes in sync.

Detect whether model predictions are already raw COCO category IDs.

Call ``_preprocess_image`` and convert NHWC -> NCHW for apples-to-apples comparison.

Write a minimal Roboflow-style COCO keypoint split.

Create a minimal YOLO segmentation dataset on disk.

Return non-zero keypoint counts from a model schema. Args: num_keypoints_per_class: Model keypoint schema. Returns: Positive

Build a ``types.SimpleNamespace`` from Pydantic model and train configs. .. deprecated:: 1.7.0 ``build_namespace`` is a backward-compatib

Build a Roboflow dataset, auto-detecting COCO or YOLO format. This function detects the dataset format and delegates to the appropriate builder f

Override PTL gradient clipping to support fused AdamW. PTL's AMP precision plugin refuses to clip gradients when the optimizer declares it ha

Seed RNGs at fit start when ``TrainConfig.seed`` is set. This avoids hidden global side-effects in ``build_trainer`` while still preserving d

Compute and display summary metrics for evaluation results.

Convert predictions to COCO format for the given iou_type.