The configuration for an RF-DETR Base model.

Track best validation mAP and save best checkpoints during training. Extends :class:`pytorch_lightning.callbacks.ModelCheckpoint` to save strippe

Validation callback that computes mAP (via torchmetrics) and macro-F1. Accumulates predictions and targets across validation batches, then at epo

Training hyperparameters and auto-batching configuration. Notes: * ``auto_batch_target_effective`` is interpreted as the **per-device**

Weight-free RFDETR that delegates to ``_DummyModel`` for all inference. Examples: >>> m = _DummyRFDETR() >>> isinstance(m.model,

LightningDataModule wrapping RF-DETR dataset construction and data loading. Args: model_config: Architecture configuration (used for reso

Wrapper to apply Albumentations transforms to (image, target) tuples. This wrapper integrates Albumentations transforms with RF-DETR's data pipel

OKS keypoint mAP metric backed by CocoEvaluator. Plain Python facade over :class:`~rfdetr.evaluation.coco_eval.CocoEvaluator` with a :meth:`r

Configuration for the preview keypoint model.

YOLO format dataset with lazy image loading and optional mask support. Both detection (``include_masks=False``) and segmentation (``include_masks

LightningModule wrapping the RF-DETR model and training loop. Args: model_config: Architecture configuration. train_config: Train

Minimal model stub that returns deterministic postprocessed results. Examples: >>> m = _DummyModel(labels=[0, 1]) >>> len(m._labe

Dataclass representing a model asset with download information. This is the standard format for model assets across rf-detr packages. Both rf-det

Batch of tensors with variable spatial sizes, padded to a common size. Stores both the padded tensor and a boolean mask indicating padding positi

Convert a raw COCO annotation dict into model-ready tensors. Accepts the ``(image, target)`` pair produced by ``torchvision.datasets.CocoDetectio

Early stopping callback monitoring validation mAP for RF-DETR. Extends :class:`pytorch_lightning.callbacks.EarlyStopping` with dual-metric mo

Training configuration for instance segmentation models. Extends :class:`TrainConfig` with segmentation-specific loss coefficients. Attribut

COCO evaluator that works in distributed mode.

Exponential Moving Average with optional tau-based warm-up. Drop-in replacement for ``rfdetr.util.utils.ModelEma`` implemented as a plain Lightni

Applies per-step drop-path and dropout rate schedules to the model. Computes the full schedule array in ``on_train_start`` using :func:`rfdetr.ut

This class computes an assignment between the targets and the predictions of the network For efficiency reasons, the targets don't include the no_

Training configuration for keypoint detection models. Extends :class:`TrainConfig` with keypoint-specific loss coefficients and metric-smooth

Minimal object supporting attribute assignment for on_load_checkpoint tests.

Minimal real nn.Module satisfying the RFDETRModule model contract. Has a single trainable parameter so the optimizer has something to update and

Callable criterion that returns a loss connected to the model output. Keeps a gradient path from the loss back to _TinyModel.dummy so that ``loss

Conv-bn module.

Dataset with ``(image, target)`` pairs for detection. The image is a ``(3, 32, 32)`` float tensor; the target dict includes the fields expected b

r"""Simplified ConvNeXt block without the MLP subnet.

Dataset wrapper that pads length to a multiple of ``effective_batch_size * world_size``. Workaround for https://github.com/Lightning-AI/pytorch-l

This is the Group DETR v3 module that performs object detection.

Convert raw RF-DETR model outputs into per-image prediction tensors. The postprocessor is shared by detection, segmentation, and keypoint inferen

The configuration for an RF-DETR Nano model.

Transformer with optional GroupPose keypoint decoder stream support.

r""" This is the configuration class to store the configuration of a [`Dinov2WithRegistersModel`]. It is used to instantiate a Dinov2WithRegis

COCO detection dataset with optional sparse-to-contiguous category ID remapping. Extends ``torchvision.datasets.CocoDetection`` with two addition

Initialize keypoint query tokens with adaptive layer-normalization style modulation.

Multi-Scale Deformable Attention Module.

Train an RF-DETR Nano model. Training accepts custom square integer ``resolution`` values. The value must be divisible by ``patch_size * num_

Train an RF-DETR Segmentation Nano model. Training accepts custom square integer ``resolution`` values. The value must be divisible by ``patch_si

This class computes the loss for Conditional DETR. The process happens in two steps: 1) we compute Hungarian assignment between ground truth

Construct the CLS token, mask token, register tokens, position and patch embeddings.

Concrete test double: provides a dropout-bearing model for eval-mode tests.

Test double for requests responses used by the downloader. Provides headers, iterable content chunks, and optional HTTP error behavior via `raise

Train or run inference with the RF-DETR keypoint preview model. Training accepts custom square integer ``resolution`` values. The value must be d

Train an RF-DETR Large model. Training accepts custom square integer ``resolution`` values. The value must be divisible by ``patch_size * num

Minimal torch.nn.Module stub shared across export tests. Avoids real forward passes; returns synthetic detection (and optionally segmentation) ou

Matcher stub used to avoid depending on Hungarian matching internals.

Minimal transformed dataset item for DataModule sample visualization.

Result of auto-batch probing: safe micro-batch size and recommended grad accumulation. Attributes: safe_micro_batch: Per-device batch siz

Utility for saving 3x3 image grids to visualize augmentation effects. Args: data_loader: Dataloader of the dataset to sample images from.

Dataclass for dataset split ratios. Attributes: train: Ratio for training set (default: 0.7) val: Ratio for validation set (defau

Very simple multi-layer perceptron (also called FFN)

Sentinel used when an augmentation entry should be removed.

Backbone shim used to validate Joiner contract changes.

Minimal nn.Module that satisfies the optimize_for_inference contract.

Tiny callable that behaves like a position encoder.

Picklable postprocessor for ddp_spawn tests. ``MagicMock`` is not picklable and cannot survive the subprocess boundary that ``ddp_spawn`` creates

Lazy YOLO dataset that defers dense mask rasterization until ``__getitem__``.

Keypoint schema inferred from COCO category metadata. Args: class_names: Category names sorted by category id. num_keypoints_per_

Very simple multi-layer perceptron (also called FFN).

This module implements MultiScaleProjector in :paper:`lwdetr`. It creates pyramid features built on top of the input feature map.

LightningCLI subclass for RF-DETR training and evaluation. Wires ``RFDETRModelModule`` and ``RFDETRDataModule`` under a unified CLI, with argumen

The configuration for an RF-DETR Large model.

The configuration for an RF-DETR Small model.

Decoder stack used by DETR transformer.

A single decoder layer with optional keypoint subnetwork.

Keypoint schema inferred from an Ultralytics YOLO pose YAML file. Args: class_names: Class names ordered by YOLO class id. num_ke

Minimal transformer surface needed for LWDETR construction and keypoint schema resizing.

Like _FakeDataset but includes binary instance masks (for segmentation).

Minimal module whose behavior differs between train and eval mode.

Minimal series object for metric grouping tests.

DDP-safe module that asserts on_train_batch_start mutation reaches training_step. With multi_scale=True and _FakeDataset's 32×32 images, on_train

Tiny LightningModule used to validate real ckpt_path resume behavior.

Simple synthetic dataset for testing augmentations and training loops. Creates synthetic images with varying numbers of bounding boxes to test ed

Minimal model that can be exported to ONNX.

backbone.

Standard bottleneck.

Faster Implementation of CSP Bottleneck with 2 convolutions.

Converts supervision Detections to the target dict format expected by RF-DETR. Args: include_masks: whether to include segmentation masks

Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).

This class turns `pixel_values` of shape `(batch_size, num_channels, height, width)` into the initial `hidden_states` (patch embeddings) of shape

The residual connection is defined in Dinov2WithRegistersLayer instead of here (as is the case with other models), due to the layernorm applied be