LibreDAMOYOLOnm / README.md
Xuban's picture
Initial upload (DAMO-YOLO-Nm, 2.7M params, before-distill, 416x416)
5d99d72 verified
---
license: apache-2.0
library_name: libreyolo
tags:
- object-detection
- damo-yolo
---
# LibreDAMOYOLOnm
DAMO-YOLO-Nm (TinyNAS_mob backbone with depthwise + GiraffeNeckV2 with
depthwise + ZeroHead with GFL/DFL, reg_max=7, last_kernel_size=1).
2.7M params, 416x416 input, repackaged for LibreYOLO.
COCO val 2017 mAP: **0.382** (per upstream README, at 416x416 input).
## Source
Derived from [tinyvision/DAMO-YOLO](https://github.com/tinyvision/DAMO-YOLO)
at commit `319572e`.
The original `damoyolo_nano_middle.pth` checkpoint was hosted on
Alibaba's Aliyun bucket which is now offline. This weight was recovered
from the [Internet Archive's Wayback Machine](https://web.archive.org/),
snapshot `20250307203300`. Upstream only ever published the
`release_model/ckpt/before_distill/` flavor of the Nano series; no
post-distill nano weights exist.
Copyright (c) 2021-2022 Alibaba Group Holding Limited. Licensed under the
Apache License, Version 2.0.
## Modifications
File rename only. `damoyolo_nano_middle.pth` is bit-identical to
`LibreDAMOYOLOnm.pt`. LibreYOLO's `nn.Module` class hierarchy mirrors
upstream's exactly so `model.load_state_dict(...)` is strict.
End-to-end bit-exact parity with upstream's pipeline verified at the
backbone, neck, and head output level (0.0 absolute difference on
identical input).
See `libreyolo/models/damoyolo/` in the
[LibreYOLO source repository](https://github.com/LibreYOLO/libreyolo)
for the architecture port.
## License
Apache License 2.0. See the [`LICENSE`](./LICENSE) and
[`NOTICE`](./NOTICE) files in this repository.