File size: 1,584 Bytes
5d99d72
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
---
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.