Return the device used by tensor outputs. Args: outputs: Model output dictionary. Top-level values are probed for tensors;

Return a new :class:`~torch.nn.Linear` resized to *num_classes* outputs. Tiles the existing weight rows when *num_classes* is larger than the cur

Apply LoRA adapters to the backbone encoder of *nn_model*. Replaces ``nn_model.backbone[0].encoder`` in-place with a PEFT-wrapped encoder using D

Build criterion and postprocessor directly from config objects. A config-native alternative to ``build_criterion_and_postprocessors(build_namespa

Pack variable-length xyxy boxes into a padded tensor and valid mask. Kornia ``AugmentationSequential`` expects boxes as ``[B, N_max, 4]``. This f

Pack variable-length instance masks into a zero-padded ``[B, N_max, H, W]`` tensor. Kornia ``AugmentationSequential`` expects masks as ``[B, N_ma

Compute many-to-many keypoint matching costs. Args: all_pred_keypoints: Predicted keypoints of shape ``(B, Q, K_total, >=7)``. ta

Deploy the trained RF-DETR model to Roboflow. Deploying with Roboflow will create a Serverless API to which you can make requests. Y

Drop scheduler.

Return a failing spec with a migration hint for removed legacy modules.

Generate a single synthetic image and its detections. Args: img_size: Side length of the square output image. min_objects: Minimu

Return the official COCO 2017 download URL for the requested asset.

Return the directory where RF-DETR caches downloaded model weights. Reads the ``RF_HOME`` environment variable; defaults to ``~/.roboflow/models`

Return True if torch.distributed is available and has been initialised.

Build a collate function that rounds batch ``H``/``W`` up to *block_size*. Used by the training DataModule to ensure that batched inputs satisfy

Apply Kornia GPU augmentation after the batch is transferred to device. When ``_kornia_pipeline`` is set and the trainer is in training mode,

A wrapper around :func:`~rfdetr.utilities.tensors._bilinear_grid_sample` to support 3D point_coords tensors. Unlike :func:`torch.nn.functional.gri

Return fixed scores/boxes (and optional keypoints) for every image in the batch.

Seed all random number generators for reproducibility. Sets seeds for Python's ``random`` module, NumPy, and PyTorch (CPU and all CUDA devices).

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

ch_in, ch_out, number, shortcut, groups, expansion.

Bilinear grid sampling compatible with all PyTorch backends including MPS. Drop-in replacement for ``F.grid_sample(input, grid, mode='bilinear',

Shared epoch-end logic for validation and test evaluation loops. Computes mAP (via ``self.map_metric``), runs the F1 confidence-threshold swe

Normalise prediction dicts from ``PostProcess`` for torchmetrics. ``PostProcess.forward`` returns masks with shape ``[K, 1, H, W]`` (the extr

Return ModelCheckpoint callbacks that are NOT BestModelCallback.

Return the :class:`~rfdetr.evaluation.keypoint_oks.MetricKeypointOKS` for *split*, creating it if needed. The metric is created lazily on fir

Return a Rich Console appropriate for the current training context. When a ``RichProgressBar`` callback is active, its ``_console`` is the Rich g

Return whether the trainer has a Lightning progress bar callback. Args: trainer: The PTL Trainer (or any object with a ``callbacks`` attr

Build ``KeyPoints`` for visualization from RF-DETR keypoint predictions. Args: key_points: RF-DETR keypoint prediction output. ke

Return a single-annotation evaluator with label2cat identity mapping.

Build a 1-image, 1-annotation, 1-keypoint COCO GT.

Build a minimal keypoint annotation used in keypoint conversion tests.

Build a bool _kp_active_mask tensor encoding *schema* (mirrors LwDetr._create_kp_active_mask). Args: schema: Keypoints-per-class list, e.

Build a minimal keypoint checkpoint encoding *kp_schema* via ``_kp_active_mask``. Args: kp_schema: Keypoints-per-class list, e.g. ``[0, 1

Build a fake nn_model mock that mimics the GroupPose keypoint model interface. Args: initial_schema: Keypoints-per-class schema the model

Create a polygon annotation from a rectangular mask region.

Build a fake PyTorch Lightning (PTL) native checkpoint with state_dict keys prefixed by 'model.'. Args: num_classes: Total classe

Create a Trainer configured for minimal smoke-test runs.

Merge a metric's accumulated per-rank state onto every rank, replacing torchmetrics' sync. torchmetrics' built-in sync (``gather_all_tensors`

Return the smallest valid config for backbone instantiation tests.

Redirect ``download_test_image_data()`` to use local calibration data. ``onnx2tf`` calls ``download_test_image_data()`` during conversion to fetc

Print overall and per-class metric tables to ``console``. Args: console: Rich ``Console`` instance to print to. title_pfx: Split

Run RFDETRCli in-process with the given args; return the SystemExit code.

Run RFDETR.train() with patched PTL; return (mock_self, output_dir path). class_names is injected via the datamodule mock (the path RFDETR.tr

Call setup('fit') with build_dataset mocked to avoid real I/O.

Decide — identically on every rank — whether to run the EMA metric ``compute()``. Under DDP, ``_merge_metric_state_across_ranks`` issues cros

Return whether the current batch closes an optimizer accumulation window. The optimizer steps when either: - The current batch close

Accumulate batch predictions into the keypoint OKS metric. Args: trainer: The PTL Trainer. outputs: Batch output dict

Write a synthetic COCO JSON file. Category IDs use sparse 1-based encoding (index * 2 + 1 → 1, 3, 5, …) so synthetic data exercises the same

Run all_gather on arbitrary picklable data (not necessarily tensors). Args: data: Any picklable object. Returns: List of dat

Build a Roboflow YOLO-format dataset. This uses Roboflow's standard YOLO directory structure (train/valid/test folders with images/ and labels/

Gather per-rank matching data from all DDP ranks and merge into one dict. Uses ``rfdetr.utilities.all_gather`` (pickle-based) so the data need no

Drop horizontal-flip transforms from keypoint augmentation configs. The helper preserves the input config shape: dictionary configs return di

Return values interpreted as non-null booleans.

Pull class names from the DataModule once the datasets are set up. Builds a ``category_id → name`` mapping from the COCO annotation metadata

Apply optional multi-scale resize to the incoming batch. Modifications to ``batch`` (in-place on ``NestedTensor``) are visible in ``training_

Prepare the EMA metric on every rank before validation (keeps DDP collectives symmetric). Args: trainer: The PTL Trainer.

Return the predict DataLoader (reuses the validation dataset, no augmentation). Returns: DataLoader for the validation dataset wi

Remove the optimized inference model and reset all optimization flags. Clears ``model.inference_model`` and resets all internal state set by

Return ``True`` after every optimizer step and every epoch end. The base ``WeightAveraging`` only updates on steps. This override also trigge

Promote a 0-D int64 scalar to a 1-D [1] int64 vector for Concat.

Return a SetCriterion with no losses so forward() is a no-op.

Build a list of ``(tensor, target_dict)`` pairs with given shapes. Args: *shapes: Variadic sequence of ``(C, H, W)`` shapes, one

Build a PTL-compatible RF-DETR checkpoint payload. Args: model_state_dict: Model weights with raw (non-prefixed) keys.

Build plot groups from numeric PTL CSVLogger metrics. Args: df: DataFrame-like object with metric columns. Returns: Non-empt

Build a single Rich renderable for overall and per-class metric tables. Args: title_pfx: Split label (e.g. ``"Val"`` or ``"Test"``).

Build minimal synthetic multi-scale inputs for `Transformer.forward`. Args: batch_size: Mini-batch size. hidden_dim: Transformer

Return a minimal model_config stub with the given patch_size.

Verify that a compatible ``onnx2tf`` package is importable. onnx2tf 2.4.0 or later is required — earlier 1.x releases cannot lower the constant `

Check whether the COCO val2017 image directory contains the expected number of JPEG files. Returns ``False`` for a missing or empty directory so

Provide a simple cross-process lock using an exclusive lock file. Args: lock_path: Path to the lock file used for mutual exclusion.

Return the set of pruneable heads and their index mask for weight pruning. Copied from transformers.pytorch_utils.find_pruneable_heads_and_indice

Convert compact GroupPose keypoints to class-padded keypoint layout.

Gather selected boxes and scale normalized coordinates to pixels. Args: out_bbox: Normalized ``cxcywh`` boxes with shape ``(B, Q,

Read the first input tensor's name and shape from an ONNX model. Args: onnx_path: Path to the ``.onnx`` file. Returns: A ``(

Build a ``(NestedTensor, targets)`` tuple for transfer_batch_to_device tests.

Create *count* small JPEG images in *directory*.

Return tensors suitable for unpack_boxes.

Build a batch with xyxy boxes suitable for on_after_batch_transfer. Returns (NestedTensor, targets) where boxes are in absolute xyxy format a

Return an RFDETR shell backed by a fully validated RFDETRBaseConfig. Uses RFDETRBaseConfig (which supplies required field defaults) so the ex

Deterministic synthetic RGB image with structure (not pure noise) so resize filtering matters.

Build a list of target dicts with the given per-image box counts. Each box is a valid xyxy rectangle within a 100x100 image.