The official five-scale Swin-L DINO checkpoint (58.5 AP on COCO val2017), packaged for LibreYOLO.

LibreYOLO ships this museum family for inference only. Its portable deformable attention uses the pure-PyTorch grid_sample reference path; the custom CUDA extension is not packaged. Native and ONNX inference use a fixed 800 x 800 PIL-bilinear stretch with ImageNet normalization. That deployment transform intentionally differs from upstream's short-side-800/max-1333 evaluation transform.

Architecture and checkpoint released by IDEA-Research/DINO at commit d84a491d41898b3befd8294d1cf2614661fc0953. Copyright (c) 2022 IDEA. All Rights Reserved. The repository declares DINO under Apache License 2.0. Its Swin-L module derives from Microsoft's MIT-licensed Swin Transformer.

The official checkpoint0027_5scale_swin.pth comes from the authors' Google Drive release. Source SHA-256: 17ddce1592816a0c63a2edc94d4a0877ffeb086f397a6657e151c703a4c850b5.

The checkpoint binary has no separate license file or license metadata. This mirror uses Apache-2.0 implied by the releasing repository's declaration, not a publisher-confirmed checkpoint-specific grant.

The complete upstream state dict is copied without renaming or transforming learned tensors. Conversion strictly loads it against the native architecture and adds LibreYOLO checkpoint metadata. The released 91-column COCO category-id head is retained and mapped to contiguous COCO-80 classes during postprocess.

The converted checkpoint SHA-256 is 1532135001dff0fa6ba688eac52df9d92af83c2c6bb13a06139fbfcd81574118. LibreYOLO's output tensors are bit-exact against the pinned upstream pure-PyTorch path (max_abs_diff == 0.0). Swin-L has near-tied, low-scoring encoder proposals that can select different background query slots across runtimes. On the real-image ONNX fixture, logit mean / p99 errors were 0.00566 / 0.0254 and box mean / p99 errors were 0.000451 / 0.00172. All 11 public detections and classes matched, minimum matched IoU was 0.9990, and maximum score difference was 0.00034.

See weights/convert_dinodetr_weights.py and docs/provenance/dinodetr.md in the LibreYOLO source repository.

Apache License 2.0 on the releasing-project implied basis described above. See the LICENSE and NOTICE files in this repository.