LibreDeformableDETRr50refine

The official iterative-box-refinement ResNet-50 Deformable DETR checkpoint (46.2 AP on COCO val2017), converted for LibreYOLO.

from libreyolo import LibreYOLO

model = LibreYOLO("LibreDeformableDETRr50refine.pt")
results = model.predict("image.jpg")

LibreYOLO ships this museum family for inference only. Its portable deformable attention uses the upstream 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.

Source

Architecture derived from fundamentalvision/Deformable-DETR at commit 11169a60c33333af00a4849f1808023eba96a931. Copyright (c) 2020 SenseTime. All Rights Reserved. Modified from DETR, Copyright 2020 - present, Facebook, Inc. Licensed under Apache License 2.0.

The checkpoint comes from SenseTime/deformable-detr-with-box-refine at revision 2e9e461623a8fdc296e19666c46c8a4389a3a6fe. Source model.safetensors SHA-256: 4113700fe8aade398808424b7c5c1304cfbf886adc6450a6ca5d50a702be3373.

Modifications

Checkpoint keys are remapped to the original module layout, decoder Q/K/V projections are concatenated, and serialization aliases for the refinement heads are restored. Non-parameter BatchNorm tracking counters are omitted; learned tensor values are unchanged. 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 84e4044553a306e07817bff9f147a50af3a51d16e43056160243a5a0d46d48c9. LibreYOLO's output tensors are bit-exact against the pinned upstream pure-PyTorch path (max_abs_diff == 0.0); fixed-800 ONNX Runtime prediction parity is also verified.

See weights/convert_deformable_detr_weights.py and docs/provenance/deformable_detr.md in the LibreYOLO source repository.

License

Apache License 2.0. See the LICENSE and NOTICE files in this repository.