license: apache-2.0
library_name: libreyolo
pipeline_tag: object-detection
tags:
- object-detection
- detr
- transformer
datasets:
- detection-datasets/coco
LibreDETRr50dc5
Original DETR-DC5 with a dilated ResNet-50 backbone (41.5M parameters, 43.3 box AP on COCO val2017 in the upstream model zoo), repackaged for LibreYOLO. DC5 replaces the final backbone stride with dilation, producing a stride-16 feature map.
from libreyolo import LibreYOLO
model = LibreYOLO("LibreDETRr50dc5.pt")
results = model.predict("image.jpg")
LibreYOLO ships this family for inference and validation, plus ONNX and TorchScript export. Training is not implemented. The deployment contract uses a fixed 800x800 canvas; upstream COCO evaluation instead preserves aspect ratio with a short side of 800 and a long side capped at 1333.
Source
Derived from the official
facebookresearch/detr checkpoint
detr-r50-dc5-f0fb7ef5.pth
at commit 29901c51d7fe8712168b8d0d64351170bc0f83e0.
Copyright (c) Facebook, Inc. and its affiliates. Licensed under the Apache
License 2.0.
Source checkpoint SHA-256:
f0fb7ef52a6b0bbd87dce4a4b8569b509f03ad082f60f936dc5b7009940295d3.
Modifications
Checkpoint metadata wrap only. Learned parameter names and tensors are
unchanged. See weights/convert_detr_weights.py in the
LibreYOLO source repository.
Strict state-dict loading succeeds with no missing or unexpected keys. Against
the pinned upstream implementation, identical input tensors produce exact
FP32 outputs (max_abs_diff == 0.0) for both pred_logits and pred_boxes.
DC5 is encoded in checkpoint metadata because it changes runtime dilation but
no parameter shape. LibreYOLO maps the sparse COCO category ids to its
contiguous 80-class public interface and does not apply NMS.
License
Apache License 2.0. See the LICENSE and NOTICE
files in this repository.