LW-DETR-tiny (12.1M params, 11.2 GFLOPs, 42.6 AP on COCO val2017), repackaged for LibreYOLO.

LW-DETR pairs a plain-ViT encoder using interleaved window/global attention with a multi-scale projector and a shallow deformable DETR decoder. It is NMS-free at inference, and it is the architecture RF-DETR was later forked from.

Inference-only in LibreYOLO: upstream's Group-DETR one-to-many training recipe is not implemented, so model.train(...) raises. Input side must be a multiple of 64 (the encoder tiles the patch grid into 4x4 windows); the default is 640.

Derived from Atten4Vis/LW-DETR (paper: arXiv 2406.03459), from the COCO checkpoint LWDETR_tiny_60e_coco.pth published at xbsu/LW-DETR. Copyright (c) 2024 Baidu. All Rights Reserved. Licensed under the Apache License 2.0.

Upstream's own lineage — Conditional DETR (Microsoft), DETR (Meta), Deformable DETR (SenseTime), and ViTDet (Meta), all Apache-2.0 — is preserved in the source headers of the LibreYOLO port.

Metadata wrap only. Learned parameters are unchanged, and no state-dict keys are renamed. See weights/convert_lwdetr_weights.py in the LibreYOLO source repository.

The upstream release was saved with --use_ema, so these are the EMA weights behind the published mAP. The 91-column classification head (one column per COCO category id) is kept as-is; LibreYOLO maps it to the contiguous 80-class interface at postprocess time.

Verified bit-exact against the official implementation: max_abs_diff == 0.0 on both pred_logits and pred_boxes.

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