---
license: mit
library_name: libreyolo
tags:
  - keypoint-detection
  - pose-estimation
  - hrnet
---

# LibreHRNetw32-pose

Official HRNet-W32 256x192 COCO-17 top-down pose weights, repackaged for
LibreYOLO.

This is a person-crop pose head. Its fixed input is one 256x192 RGB person
crop; the checkpoint does not contain a person detector. LibreYOLO can compose
the head with its default YOLO9 detector, another `PersonDetector`, explicit
`person_boxes`, or `cropped=True`.

## Source

Derived from
[`leoxiaobin/deep-high-resolution-net.pytorch`](https://github.com/leoxiaobin/deep-high-resolution-net.pytorch)
at commit `6f69e4676ad8d43d0d61b64b1b9726f0c369e7b1`.
The upstream repository license is Copyright (c) 2019 Leo Xiao and is MIT.
Adapted implementation files also identify Microsoft and Bin Xiao.

Official source file: `pose_hrnet_w32_256x192.pth`

Source SHA-256:
`19bc083708bb8d873211e50d85d56344c10290c6e8b564c813fdde09645c4c1c`

## Modifications

The original state dict is wrapped with LibreYOLO checkpoint metadata. Learned
tensor keys, values, and dtypes are unchanged and strict-load into the native
HRNet-W32 graph. The converted file is `LibreHRNetw32-pose.pt`, SHA-256
`c9d8eb383e63ce795f87a3ee2f2b0c8aa0df9bae981573ec4b3909d63815a825`.
See `weights/convert_hrnet_weights.py` and `docs/provenance/hrnet.md` in the
[LibreYOLO source repository](https://github.com/LibreYOLO/libreyolo).

## License

The official project distributes this checkpoint from its MIT-licensed model
zoo but does not attach a separate checkpoint-specific license. This mirror's
redistribution basis is the MIT license implied by the releasing project, not
a claim of an independently confirmed per-file grant. Training-data rights
remain separate. See [`LICENSE`](./LICENSE) and [`NOTICE`](./NOTICE).