CondadosAI commited on
Commit
f7bf367
·
verified ·
1 Parent(s): 9186624

docs: add NOTICE attribution per Apache-2.0 s4(d) (code AND weights)

Browse files
Files changed (1) hide show
  1. NOTICE +61 -0
NOTICE ADDED
@@ -0,0 +1,61 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ RTMO-s (body7) — acaua mirror (pure-PyTorch port)
2
+ ====================================================
3
+
4
+ This product includes:
5
+
6
+ 1. PORTED SOURCE CODE: a pure-PyTorch port of the RTMO architecture
7
+ (located in the acaua repository at src/acaua/adapters/rtmo/) is a
8
+ derivative work of:
9
+
10
+ - OpenMMLab's mmpose implementation
11
+ https://github.com/open-mmlab/mmpose @ commit
12
+ 759b39c13fea6ba094afc1fa932f51dc1b11cbf9 — Apache-2.0
13
+ Files derived from:
14
+ mmpose/models/backbones/csp_darknet.py
15
+ mmpose/models/necks/hybrid_encoder.py
16
+ mmpose/models/heads/hybrid_heads/rtmo_head.py
17
+ mmpose/models/heads/hybrid_heads/yoloxpose_head.py
18
+ mmpose/models/utils/csp_layer.py
19
+ mmpose/models/utils/reparam_layers.py
20
+ mmpose/models/utils/transformer.py
21
+ mmpose/evaluation/functional/nms.py
22
+ - Plus mmcv primitives (ConvModule, DepthwiseSeparableConvModule,
23
+ FFN, MultiheadAttention, Scale) vendored as pure-PyTorch
24
+ equivalents.
25
+
26
+ Paper: Peng Lu, Tao Jiang, Yining Li, Xiangtai Li, Kai Chen,
27
+ Wenming Yang, "RTMO: Towards High-Performance One-Stage Real-Time
28
+ Multi-Person Pose Estimation", CVPR 2024 (arXiv:2312.07526).
29
+
30
+ 2. CONVERTED WEIGHTS: the model.safetensors file in this mirror is a
31
+ key-remapped conversion of the upstream pretrained checkpoint:
32
+
33
+ - upstream URL: https://download.openmmlab.com/mmpose/v1/projects/rtmo/rtmo-s_8xb32-600e_body7-640x640-dac2bf74_20231211.pth
34
+ - upstream SHA256: dac2bf749bbfb51e69ca577ca0327dff4433e3be9a56b782f0b7ef94fb45247e
35
+ - upstream paper: Lu et al., CVPR 2024 (arXiv:2312.07526)
36
+ - training set: body7 = COCO + AI Challenger + CrowdPose + MPII
37
+ + sub-JHMDB + Halpe + PoseTrack18
38
+
39
+ Conversion was performed by scripts/convert_rtmo.py in the acaua
40
+ repository. The conversion is deterministic and reversible: state
41
+ dict keys can be remapped back and are lossless (no quantization
42
+ or pruning applied). Training-only loss-module buffers
43
+ (head.loss_oks.sigmas, etc.) and data-preprocessor buffers
44
+ (data_preprocessor.mean, .std) are stripped — they have no role
45
+ at inference and are regenerated from config at training time.
46
+
47
+ Mirrored on 2026-04-22 by CondadosAI.
48
+
49
+ License
50
+ -------
51
+ Licensed under the Apache License, Version 2.0 (the "License");
52
+ you may not use this file except in compliance with the License.
53
+ You may obtain a copy of the License at
54
+
55
+ http://www.apache.org/licenses/LICENSE-2.0
56
+
57
+ Unless required by applicable law or agreed to in writing, software
58
+ distributed under the License is distributed on an "AS IS" BASIS,
59
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
60
+ implied. See the License for the specific language governing
61
+ permissions and limitations under the License.