obb_loss

Typed PyTorch oriented-box losses and standalone matching for DETR-like detectors. The package exposes pure functions, Pydantic-configured nn.Module criteria, and a typed custom-function builder. Detector-layer orchestration stays in your pipeline.

See runnable examples for complete CPU forward/backward scripts.

The primary geometry is (cx, cy, width, height, cos(theta), sin(theta)), with positive sides and nonzero direction vectors. Coordinate units are explicit and shared by predictions and targets. Use boxes_from_angles to convert radians; no format inference occurs.

Pack existing tensors

import obb_loss as obb
import torch

# Existing per-sample target tensors, including an empty sample.
targets_result = obb.pack_boxes(
    [torch.tensor([[0., 0., 2., 1., 1., 0.]]), torch.empty(0, 6)],
    device=torch.device("cpu"),
    dtype=torch.float32,
)
# Dense detector output [batch, queries, 6] also has a helper.
predictions = torch.tensor([[[0.1, 0., 2., 1., 1., 0.]],
                            [[0.2, 0., 2., 1., 1., 0.]]], requires_grad=True)
preds_result = obb.pack_dense_boxes(predictions)

Helpers preserve gradients and sample boundaries, never mutate or silently transfer inputs, and return Ok(PackedBoxes) or Err(InputError). Device and dtype remain explicit even for zero-sample input. PackedBoxes.boxes is [total_boxes, 6]; cu_seqlens is integer [batch + 1], starts at zero, ends at total_boxes, and retains empty samples through repeats.

Match, then evaluate an aligned box loss

import obb_loss as obb

match (preds_result, targets_result):
    case (obb.Ok(value=preds), obb.Ok(value=targets)):
        matcher = obb.HungarianMatcher(obb.MatcherConfig(chunk_size=4096, max_pairs=1_000_000))
        match matcher.forward(preds, targets):
            case obb.Ok(value=assignment):
                matched_preds, matched_targets = obb.gather_matched(preds, targets, assignment)
                criterion = obb.ProbIoULoss(obb.ProbIoUConfig(reduction=obb.Reduction.MEAN))
                loss = criterion.forward(matched_preds, matched_targets)
                loss.backward()
            case obb.Err(error=error):
                print(error)
    case _:
        print("Packing failed; inspect the typed error outcome")

Box criterion signature: forward(preds: PackedBoxes, targets: PackedBoxes) -> Tensor. forward trusts geometry, aligned counts, and offsets; checked validates fresh inputs and returns a typed result. Pure stable_probiou_loss returns unreduced values and stable_probiou returns similarity. Outputs are float64; autocast is disabled locally without changing your detector's outer mixed-precision context. Empty mean/sum is a gradient-connected zero. Mean averages matched pairs across the whole pack, not per-sample means.

The matcher solves each sample independently, assigns the smaller side fully, and returns detached indices on the input device. Its CPU solver makes an explicit cost transfer from CUDA. Cost construction is chunked; complete cost storage remains quadratic within each sample. See matching contracts for limits and solver references.

Quality classification

For one complete decoder layer, obb.SetCriterion owns matching and weighted classification, L1, and ProbIoU losses. Use obb.pack_predictions(boxes, logits) and obb.pack_targets(samples, device=..., dtype=...), where each obb.TargetSample contains boxes, class labels, and caller-supplied quality. Match obb.Ok(value=losses) from criterion.forward(preds, targets) and backpropagate losses.total. obb.SetCriterionConfig.weights controls training; .matcher.weights controls assignment independently. See the DETR example and set criterion contracts.

import obb_loss as obb

logits = torch.tensor([[0.3, -0.8]], requires_grad=True)
quality = torch.tensor([[0.7, 0.]])  # Caller-supplied [queries, classes], zero = negative.
vfl = obb.VarifocalLoss(obb.VFLConfig(alpha=0.75, gamma=2., quality_weighted=True))
qfl = obb.QualityFocalLoss(obb.QFLConfig(beta=2.))
classification_loss = vfl.forward(logits, quality)  # Or qfl.forward(logits, quality).

Pure varifocal_loss and quality_focal_loss return unreduced [queries, classes] values. Checked modules reject nonfinite logits and quality outside [0, 1]. Targets are not automatically detached; choose that policy explicitly. Mean averages all class elements; positive-count or distributed normalization is yours. No auxiliary decoder-layer loop or classification target synthesis is included. See loss papers and numerical policy.

Custom functions and typed settings

import obb_loss as obb
from torch import Tensor

def center_loss(preds: Tensor, targets: Tensor) -> Tensor:
    """Return an unreduced center penalty for already aligned boxes."""
    return (preds[:, :2] - targets[:, :2]).square().sum(-1)

builder = (obb.BoxLossBuilder()
           .add("probiou", obb.stable_probiou_loss)
           .add("center", center_loss, weight=0.1))
custom_result = builder.build(obb.LossConfig(reduction=obb.Reduction.MEAN))
config_result = obb.parse_config(obb.ProbIoUConfig, {"weight": 2., "reduction": "sum"})

Functions receive float64 aligned tensors and return finite same-device [pairs] losses. The library applies weights/reduction once. checked catches observable output-contract failures; user code bugs propagate. nn.Module components are registered as children, including parameters and buffers, so .to() and state_dict() work. Zero-weight terms are omitted. Builders are immutable descriptions; caller-owned modules can still be shared if deliberately supplied to multiple builders. Use the same narrow function contract for custom matcher costs. Config JSON cannot serialize/reconstruct Python callbacks; recreate custom composition before loading weights.

parse_config accepts JSON-compatible values and returns typed validation errors for external settings. Direct Pydantic model construction retains its ordinary exception semantics. build_loss constructs a built-in module from validated ProbIoUConfig, VFLConfig, or QFLConfig.

Verification

Use uv sync to install the development environment and just check to run strict formatting, lint, type, and test checks. Python development is pinned to 3.11; package metadata supports Python 3.11+. Tool settings live separately in ruff.toml, pyrefly.toml, and pytest.ini. Each check recipe synchronizes its environment.

just stress --samples 10000 --chunk-size 512 --dtype FP64
just benchmark --samples 4 --queries 128 --targets 32
just test-cuda
just build

Stress and benchmark commands emit JSON on stdout and progress on stderr; use --quiet to suppress progress. test-cuda fails if hardware is unavailable, so skipped cases cannot masquerade as GPU verification. Read verification evidence for measured results and limitations. Build creates a portable pure-Python wheel; nothing publishes.