Infer keypoint count for one category from annotation vectors. Args: annotations: COCO annotation dictionaries. category_id: Cate

Compute log mean covariance trace for active keypoints. Args: active_keypoints: Active keypoint predictions with shape

Build flat keypoint prediction rows for notebook or terminal display. Args: key_points: RF-DETR keypoint prediction output. image

Convert keypoints to numpy array and validate shape. >>> import torch >>> keypoints = torch.tensor([[[10.0, 20.0, 2.0]]]) >>>

Return the plotting line style for a metric split.

List YOLO image files in a stable order.

Load cookbook card definitions from a YAML file. Reads the YAML file and returns the ``cards`` list, which is consumed by the cookbook landin

Load a COCO annotation JSON file. Args: annotation_path: Path to a COCO annotation JSON file. Returns: Parsed COCO annotatio

Load project metadata from a pyproject file. Reads the TOML file in binary mode, matching the parser API. The returned dictionary is used by

Log compact per-step convergence metrics for the progress bar only. Args: loss: Unscaled aggregate training loss. los

Log aggregate and component validation losses. Args: loss: Aggregate weighted validation loss. loss_dict: Raw criteri

Build a batch with xyxy boxes and instance masks for segmentation tests. Returns (NestedTensor, targets) where each target includes a 'masks'

Build a minimal legacy .pth checkpoint (no ``state_dict`` key). Args: pe_size_src: Source grid side length. dim: Embe

Build a minimal (samples, targets) batch for probing. Uses max_targets_per_image targets per image so memory reflects worst-case matcher and loss

Return a minimal TrainConfig for use in load_pretrain_weights. Returns: Minimal TrainConfig with placeholder dataset and output dirs.

Build an in-memory ZIP archive from a mapping of filename→content.

Return element-wise maximums of a list of lists. Args: the_list: List of integer lists, all of the same length. Returns: Lis

Return whether a torchmetrics metric has accumulated at least one update.

Normalize transform params across Albumentations API variations. Currently this adapts ``RandomSizedCrop`` arguments so a config using ``height``

Normalize a keypoint name for symmetry matching.

Normalize split ratios parameter to a dictionary. Args: split_ratios: Can be: - DatasetSplitRatios dataclass instance

ONNX-tracing-compatible variant of ``nested_tensor_from_tensor_list``. Args: tensor_list: List of 3-D tensors (C, H, W). block_si

Parse a YOLO center-width-height box into relative XYXY coordinates.

Parse a flattened YOLO polygon into relative XY points.

Parse one Ultralytics YOLO pose row into pixel boxes and COCO-style keypoints.

Return a minimal COCO-like object with a single keypoint-bearing person category.

Rasterize a polygon into a dense boolean mask. TODO: remove once supervision ships a direct CompactMask.from_polygon factory; at that point t

Rasterize per-instance polygons into an ``(N, H, W)`` boolean array. TODO: remove once supervision ships a direct CompactMask.from_polygon factor

Build detection-only result dictionaries. Args: scores: Selected object scores with shape ``(B, K)``. labels: Selecte

Select class-specific keypoints and optionally fuse object scores. Args: out_keypoints: Raw keypoint predictions with shape

Attach resized segmentation masks for selected detections. Args: out_masks: Raw mask logits with shape ``(B, Q, Hm, Wm)``.

Single detection prediction dict for image_id.

Build a dynamic-range INT8 TFLite model from the onnx2tf SavedModel. Dynamic-range quantization stores weights as INT8 and keeps activations in f

Return whether ``RandomSizedCrop`` expects a ``size`` keyword. The Albumentations 2.x API changed ``RandomSizedCrop`` from separate ``height``/``

Read the package version from pyproject project metadata. Validates that the version exists and is a string before exposing it to the documen

Rebuild transformed keypoints and keep them synchronized with kept boxes. Args: augmented: Augmented output dict from Albumentati

Rewrite one GridSample ONNX node into a TFLite-safe bilinear subgraph. Replaces ``GridSample(im, grid)`` with an equivalent bilinear sampling sub

Return whether the configured devices value explicitly requests multiple devices.

Resolve OKS sigmas for one keypoint-count group.

Resolve OKS sigmas for faster-coco-eval keypoint evaluation.

Map a torch-style device specifier to PTL ``accelerator``/``devices`` kwargs. Args: device: A device specifier accepted by ``torc

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

Return a copy of *predictions* with all tensors detached and moved to CPU. Prevents callers from inadvertently retaining CUDA memory or autograd

Write a small solid-colour grayscale PNG to *path*.

Scale the current numerator loss by the accumulated box denominator. Args: raw_loss: Current microbatch weighted loss numerator.

Select the highest scoring query/class pairs. Args: out_logits: Classification logits with shape ``(B, Q, C)``. Returns:

Call setup('fit') with build_dataset and cuda mocked (no CUDA → fallback).

Resolve augmentation backend and build the Kornia pipeline if applicable. Called once during ``setup("fit")``. When ``augmentation_backend``

Return whether initialization should retry with deprecated Large config. The fallback is only for known checkpoint/config incompatibilities f

Return a source image path for common COCO-style datasets.

Split a CSVLogger metric column into split prefix and metric name.

Step Lightning's scheduler object when one is configured.

Clip gradients, step optimizer and scheduler, then reset accumulation state. Args: optimizer: Optimizer returned by Lightning.

Verify predict() uses sparse COCO category-ID → class-name mapping. Issue #988: RFDETRSegSmall returned "sheep" for class_id=18 instead of "dog"

Round-trip a model through from_checkpoint using a temp training checkpoint. Saves the instantiated model's weights into a minimal training-style

Probe the full tensorboard import chain to surface numpy/tensorflow incompatibilities early. When tensorboard is installed alongside a numpy-2.0-

Validate COCO annotations container type. Args: annotations: Raw ``annotations`` value from a COCO annotation file. Returns:

Validate and sort COCO categories by category id. Args: categories: Raw ``categories`` value from a COCO annotation file. Returns:

Validate mutually exclusive output heads and per-image target sizes. Args: out_logits: Classification logits with shape ``(B, Q,

Validate and normalize a YOLO pose ``kpt_shape`` entry.

Emit the standard warning for a disabled keypoint horizontal flip.

Warn once when metric table rendering is skipped because Rich is unavailable. Args: warning_emitted: Whether this warning has already bee

Compute category-weighted mean COCO stats, ignoring unavailable ``-1`` values.

Write a COCO keypoint file with two categories using different keypoint counts.

Computes the precision@k for the specified values of k.

Register argument links that share configs between module and datamodule. Linking ``model.model_config`` → ``data.model_config`` and ``model.

Compute the DICE loss, similar to generalized IOU for masks. Args: inputs: A float tensor of arbitrary shape. The predictions for each ex

Compute sigmoid cross-entropy loss for mask predictions. Args: inputs: A float tensor of arbitrary shape. The predictions for each exampl

Useful args: - encoder: encoder name - lr_encoder: - dilation - use_checkpoint: for swin only for now

Build a HungarianMatcher from a training argument namespace. Args: args: Namespace supplying ``focal_alpha``, ``set_cost_class``, ``set_c

Build a Kornia ``Normalize`` transform for GPU-side normalization. Args: mean: Per-channel mean values. Defaults to ImageNet statistics.

We estimate uncertainty as L1 distance between 0.0 and the logit prediction in 'logits' for the foreground class in `classes`. Args:

Detect if a Roboflow dataset is in COCO or YOLO format. Args: dataset_dir: Path to the Roboflow dataset root directory Returns:

Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks). Comment by Ross Wightman: This is the same as the DropCo

Expand and resolve the pretrain_weights path. Bare filenames (no directory component, e.g. ``rf-detr-base.pth``) are resolved to the model ca

Export the trained model to ONNX or TFLite format. See the `export documentation <https://rfdetr.roboflow.com/learn/export/>`_ for more infor

Generalized IoU from https://giou.stanford.edu/ The boxes should be in [x0, y0, x1, y1] format. Returns a [N, M] pairwise matrix, where N =

Get activation.

Build binddings.

Calculate lr decay rate for different ViT blocks. Args: name: Parameter name. lr_decay_rate: Base lr decay rate. num_laye