| ---
|
| license: apache-2.0
|
| library_name: libreyolo
|
| pipeline_tag: object-detection
|
| tags:
|
| - object-detection
|
| - lw-detr
|
| - detr
|
| datasets:
|
| - detection-datasets/coco
|
| ---
|
|
|
| # LibreLWDETRm
|
|
|
| LW-DETR-medium (28.2M params, 42.8 GFLOPs, 52.5 AP on COCO val2017), repackaged for [LibreYOLO](https://github.com/LibreYOLO/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.
|
|
|
| ```python
|
| from libreyolo import LibreYOLO
|
|
|
| model = LibreYOLO("LibreLWDETRm.pt") # downloads from this repo
|
| results = model.predict("image.jpg")
|
| ```
|
|
|
| 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.
|
|
|
| ## Source
|
|
|
| Derived from [Atten4Vis/LW-DETR](https://github.com/Atten4Vis/LW-DETR)
|
| (paper: [arXiv 2406.03459](https://arxiv.org/abs/2406.03459)), from the COCO
|
| checkpoint `LWDETR_medium_60e_coco.pth` published at
|
| [xbsu/LW-DETR](https://huggingface.co/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.
|
|
|
| ## Modifications
|
|
|
| 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](https://github.com/LibreYOLO/libreyolo).
|
|
|
| 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`.
|
|
|
| ## License
|
|
|
| Apache License 2.0. See the [`LICENSE`](./LICENSE) and [`NOTICE`](./NOTICE)
|
| files in this repository.
|
|
|
