Input
: [[1, 3, 256, 192]] — RGB, ImageNet norm, 256x192 person crop (detect first, then crop and resize to this aspect)
Output
: SimCC pair: x [1,17,384] and y [1,17,512] — a 1-D distribution per keypoint per axis. Decode: keypoint k sits at (argmax(x[k]) / 2, argmax(y[k]) / 2) in crop pixels; the max value doubles as the confidence.
Variants
All variants take and return fp32 tensors — swap the
.pte
file, keep your app code.
precision
file
size (MB)
parity vs fp32 eager (worst corr)
Mac median (ms)*
fp32
rtmpose_s_body_xnnpack_fp32.pte
21.9
1.000000
5.7
*Mac arm64, single process, median of 10 — a reference point for relative cost
only, not a device number (torch eager fp32 on the same machine: 34.3 ms).
Precisions that did not earn a slot
fp16 is not shipped
: keypoints move a median of 16 px and up to 41 px on a 192x256 crop (169 confident keypoints across 12 person crops). It also only saves 5% of the file, so there is nothing to weigh against that.
int8 is not shipped
: keypoints move a median of 131 px on a 192x256 crop — the pose is gone. CSPNeXt's channel attention and the head's gated attention unit both resist post-training quantization; this needs QAT, not a different calibration set (re-calibrating on person crops rather than scenes made it worse, not better).
Verification (executorch 1.4.0, torch 2.13.0)
Parity is measured against the fp32 eager model on real image input;
corr
is
the correlation over all elements of each output tensor.
output
shape
max_abs_diff
corr
0
[1, 17, 384]
6.855e-06
1.000000
1
[1, 17, 512]
8.076e-06
1.000000
XNNPACK delegate coverage (fp32): 92.0% (229/249 ops); ops left on the portable kernels:
dim_order_ops._to_dim_order_copy.default
x8,
aten.unsqueeze_copy.default
x3,
aten.split_with_sizes_copy.default
x3,
aten.sum.dim_IntList
x2,
aten.pow.Tensor_Scalar
x2,
aten.squeeze_copy.dims
x2
RTMPose-s-Body-ExecuTorch huggingface.co is an AI model on huggingface.co that provides RTMPose-s-Body-ExecuTorch's model effect (), which can be used instantly with this mlboydaisuke RTMPose-s-Body-ExecuTorch model. huggingface.co supports a free trial of the RTMPose-s-Body-ExecuTorch model, and also provides paid use of the RTMPose-s-Body-ExecuTorch. Support call RTMPose-s-Body-ExecuTorch model through api, including Node.js, Python, http.
RTMPose-s-Body-ExecuTorch huggingface.co is an online trial and call api platform, which integrates RTMPose-s-Body-ExecuTorch's modeling effects, including api services, and provides a free online trial of RTMPose-s-Body-ExecuTorch, you can try RTMPose-s-Body-ExecuTorch online for free by clicking the link below.
mlboydaisuke RTMPose-s-Body-ExecuTorch online free url in huggingface.co:
RTMPose-s-Body-ExecuTorch is an open source model from GitHub that offers a free installation service, and any user can find RTMPose-s-Body-ExecuTorch on GitHub to install. At the same time, huggingface.co provides the effect of RTMPose-s-Body-ExecuTorch install, users can directly use RTMPose-s-Body-ExecuTorch installed effect in huggingface.co for debugging and trial. It also supports api for free installation.
RTMPose-s-Body-ExecuTorch install url in huggingface.co: