Move tensors and mask to *device*. Args: device: Target device. **kwargs: Additional arguments forwarded to ``Tensor.to``.

Assemble a PTL ``Trainer`` with the full RF-DETR callback and logger stack. Resolves training precision from ``model_config.amp`` and device capa

Create a minimal mock Trainer with controllable callback_metrics. Sets the attributes required by ModelCheckpoint and EarlyStopping skip-guards s

Return a minimal mock Trainer with an optional DataModule.

Train an RF-DETR model via the PyTorch Lightning stack. All keyword arguments are forwarded to :meth:`get_train_config` to build a :class:`~r

Create a minimal mock RFDETRModule with state_dict and train_config.

Return a minimal mock LightningModule.

Minimal RFDETRBaseConfig for tests.

Save best regular/EMA checkpoints when validation mAP improves. Delegates regular-model checkpoint management to the :class:`~pytorch_lightni

Minimal TrainConfig for tests. Loggers are disabled by default to avoid requiring optional deps (tensorboard, wandb, mlflow) in the CPU test

Performs model inference on the input images. This method accepts a single image or a list of images in various formats (file path, image url

Load pretrained checkpoint weights into *nn_model* in-place. Canonical implementation shared by the L1 facade (``_build_model_context`` in ``rfde

Return a MagicMock shaped like RFDETR with real config objects. No spec is used because RFDETR.model is set in __init__ (instance attr) and spec=

No-op progress update for compatibility with tqdm.

Factory fixture — call with **overrides to get a minimal RFDETRBaseConfig.

Instantiate ``MeanAveragePrecision`` after DDP device placement. Args: trainer: The PTL Trainer. pl_module: The Light

Creates and configures a logger with stdout and stderr handlers. This function creates a logger that sends INFO and DEBUG level logs to stdout, a

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

Build a list of :class:`AlbumentationsWrapper` instances from a config. Supports both a flat dictionary format (backward-compatible) and a li

Build a ``types.SimpleNamespace`` from configs and architectural defaults. This is the internal implementation behind :func:`build_namespace`. Ex

Return whether any value is truthy.

Optimize the model for inference with optional JIT compilation and dtype casting. Operations are wrapped in the correct CUDA device context t

Return a minimal ``_FakeDataset`` with a controllable length.

Factory fixture — returns (module, fake_model, fake_criterion, fake_postprocess). build_model and build_criterion_and_postprocessors are mocked a

Build a namespace suitable for builder functions.

Compute effective mAP and delegate to parent stopping logic. Computes ``ema_mAP`` or ``max(regular_mAP, ema_mAP)`` depending on ``use_ema``,

Return a MagicMock that behaves enough like an LWDETR nn.Module. Returns: MagicMock with reinitialize_detection_head and load_state_dict

Construct RFDETRModelModule with build_model_from_config and build_criterion_from_config mocked.

Return a minimal torchmetrics-style metrics dict.

Factory fixture — returns a constructed RFDETRDataModule. build_dataset is mocked automatically. tmp_path is injected automatically so test metho

Return callback state for checkpointing.

Run OKS keypoint evaluation and return metric dict. Constructs a fresh :class:`~rfdetr.evaluation.coco_eval.CocoEvaluator`, replays a

Download pretrained weights with optional MD5 validation. Download Priority Order: The function searches for models in the following orde

Invoke RFDETR.from_checkpoint with torch.load mocked to return *ckpt* and the model class at *cls_patch_target* replaced by a MagicMock. Retu

Preprocess one image, run ONNX Runtime inference, and decode detections. Reads input shape from the session (NCHW ``float32``), resizes and norma

Convert an ONNX model to TFLite via ``onnx2tf``. Requires ``onnx2tf >= 2.4.0``. Uses the Python API with a NumPy compatibility shim. Args:

Accumulate per-batch predictions. Each call appends one batch; predictions are replayed in order inside :meth:`compute`. Predictions

Return a minimal RFDETRBaseConfig with pretrain_weights disabled.

Load an RF-DETR model from a training checkpoint, automatically inferring the model class. The correct subclass is resolved in order of prefe

Build the training resize pipeline as an Albumentations config list. Expresses the ``RandomSelect(resize_a, Compose([resize_b1, crop, resize_b2])

Build a (NestedTensor, targets) tuple for testing.

Build datasets for the requested stage. PTL calls this on every process before the corresponding dataloader method. Datasets are built lazil

Return a minimal RFDETRBaseConfig with pretrain_weights disabled.

Return a minimal TrainConfig suitable for unit tests.

Return a minimal TrainConfig suitable for unit tests.

Instantiate ``config_cls(**kwargs)`` and return only the pretrain-compat warnings.

Return a minimal COCO ground-truth mock (for unit tests that patch evaluator).

Infer a keypoint schema from a COCO annotation JSON file. Args: annotation_path: Path to a COCO annotation JSON file. keypoint_ok

Compute gradients with cuDNN disabled. Args: ctx: Autograd context with saved tensors and conv parameters. grad_outpu

Auto-detect legacy formats and reconcile PE shapes at checkpoint load time. PTL calls this hook before applying ``checkpoint["state_dict"]``

Accumulate per-image predictions.

Validate that a checkpoint is compatible with the model configuration. Checks for mismatches in ``segmentation_head`` and ``patch_size`` between

Return a compact matching dict as produced by ``build_matching_data()``.

Build compact per-class matching data from a batch of predictions and targets. Implements greedy highest-score-first matching compatible with the

Convert a legacy RF-DETR ``.pth`` checkpoint to PTL ``.ckpt`` format. Loads a checkpoint saved by the pre-PTL ``engine.py`` training loop and rew

Return callback state including ``_best_ema``, ``_smoothed_regular``, and ``_best_raw_regular``. Extends the parent :class:`~pytorch_lightnin

Force the manual gather-based code path by mocking input.device.type. The function checks ``input.device.type != "mps"`` to decide which branch t

Ground-truth output from F.grid_sample for comparison.

Build a mock TFLite interpreter with configurable I/O details.

Validate a user-supplied ``(height, width)`` shape tuple and return normalised plain-int dims. Args: shape: The raw value supplied by the

Build a Kornia ``AugmentationSequential`` from an aug_config dict. Each key in *aug_config* is looked up in ``_REGISTRY`` and instantiated with t

Merge image IDs and COCO result records across distributed processes. Each image ID is assigned to exactly one rank (first rank that reports

Build a query embedding tensor where row ``g * num_queries + q`` encodes ``[g * 100 + q, 0, ...]``. This lets tests check the per-group ordering

Merge *new_data* into *accumulator* in place. Both arguments share the dict schema produced by ``build_matching_data()``: each class-keyed sub-di

Assert resize from on_train_batch_start propagated before calling super.

Return (1, 10, 82) logits with one high-confidence entry when requested. Background fill is -10.0 so sigmoid scores are near zero (~0.0001) for a

Convert COCO segmentation annotations to a binary mask tensor of shape ``[N, H, W]``. Supports both polygon and RLE (Run-Length Encoding) annotat

Select the overall best model and optionally run test evaluation. Copies the winner (regular vs EMA, strict ``>`` for EMA) to ``checkpoint_be

Build a ModelContext from ModelConfig without using legacy main.py:Model. Replicates ``Model.__init__`` logic: builds the nn.Module, optionally l

Emit a DeprecationWarning pointing users to the new module location. Args: old: Fully-qualified name of the deprecated module (e.g. ``rfd

Accumulate per-image evaluation results into mean metrics.

Return an empty matching accumulator compatible with ``merge_matching_data()``. Returns: Empty dict to be passed as the first argument to

Restore callback state from checkpoints.

Restore callback state from a Lightning checkpoint. Pops ``"_best_ema"``, ``"_smoothed_regular"``, and ``"_best_raw_regular"`` from a shallow

Build the standard COCO transform pipeline for a given dataset split. Returns a composed transform that resizes images to the target ``resolution

Create COCO transforms with square resizing where the output size is divisible by 64. This function builds a torchvision-style transform pipeline

Check whether a COCO segmentation entry is in RLE format. RLE annotations are dicts with ``"counts"`` and ``"size"`` keys, as opposed to polygon

Minimal KeypointTrainConfig for tests that exercise keypoint model paths.

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