File size: 3,060 Bytes
be7affb | 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 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 | UniFormer-S COCO top-down pose - acaua mirror (pure-PyTorch port)
==================================================================
This product includes:
1. PORTED SOURCE CODE: a pure-PyTorch port of the UniFormer-pose
architecture (located in the acaua repository at
src/acaua/adapters/uniformer/pose/) is a derivative work of:
- Sense-X/UniFormer (Apache-2.0)
https://github.com/Sense-X/UniFormer @ main
Files derived from:
pose_estimation/mmpose/models/backbones/uniformer.py
(the global-attention UniFormer backbone, parameter-tree
identical to the detection variant — we reuse acaua's
UniFormer2DDense with hybrid=False, windows=False)
pose_estimation/mmpose/models/keypoint_heads/top_down_simple_head.py
(TopDownSimpleHead: 3x ConvTranspose2d-stride-2 + BN + ReLU
+ 1x1 final conv)
pose_estimation/mmpose/core/evaluation/top_down_eval.py
(keypoints_from_heatmaps with post_process='default' branch:
argmax + sign(neighbor_diff) * 0.25 shift)
2. CONVERTED WEIGHTS: the model.safetensors file in this mirror is a
key-remapped conversion of the upstream pretrained checkpoint:
- upstream repo: Sense-X/UniFormer (Apache-2.0)
- upstream weights: Google Drive file id
162R0JuTpf3gpLe1IK6oxRoQK7JSj4ylx
- upstream filename: top_down_256x192_global_small.pth
- upstream SHA256: d77059e3e9322c0e20dc89dc0cf2a583ffe2ced7d3e9b350233738add570bc30
- upstream paper: Li et al., "UniFormer: Unifying Convolution
and Self-attention for Visual Recognition",
arXiv:2201.09450 (ICLR 2022)
Key remap rule (eager, at conversion time):
backbone.* -> backbone.* (no change)
keypoint_head.* -> head.*
Conversion was performed by scripts/convert_uniformer_pose.py in the
acaua repository. The adapter loads the safetensors file under
load_state_dict(strict=True) — key-naming drift between upstream and
the port is a hard error at load time.
3. BUNDLED DETECTOR: this mirror's adapter loads
CondadosAI/rtmdet_t_coco at adapter-init time so callers get the
full top-down pipeline (detect persons -> per-person crop -> pose ->
inverse-warp) from a single predict(image) call. RTMDet-tiny carries
its own attribution chain — see CondadosAI/rtmdet_t_coco's NOTICE
file.
Mirrored on 2026-04-25 by CondadosAI.
License
-------
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
|