mlboydaisuke/RTMPose-m-Hand-ExecuTorch

rtmpose_m_hand — ExecuTorch

• Source: open-mmlab/mmpose RTMPose (rtmpose-m_simcc-hand5_pt-aic-coco_210e-256x256-74fb594)
• License: Apache-2.0
• Input: [[1, 3, 256, 256]] — RGB, ImageNet norm, 256x256 crop around a hand (detect first, then crop and resize to this aspect)
• Output: SimCC pair: x [1,21,512] and y [1,21,512] — 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.

build	file	size (MB)	parity vs fp32 eager (worst corr)	Mac median (ms)*
fp32	rtmpose_m_hand_xnnpack_fp32.pte	55.1	1.000000	9.5
Core ML (fp16, iOS)	rtmpose_m_hand_coreml_all.pte	28.0	0.999768	2.9

The Core ML build is the same graph lowered to Apple's Neural Engine instead of XNNPACK, which is CPU-only. Measured on an iPhone 17 Pro across seven models, it runs 3.5x to 13.9x faster (median 12x) at roughly half the file size — for example Depth-Anything-V2-Small at 500.8 ms against 42.7 ms, and MODNet at 81.7 ms against 5.9 ms. It computes in fp16 and is iOS-only; the XNNPACK files stay the portable option and are what runs on Android.

*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: 90.7 ms).

Checked in the task's own units

Correlation is a first filter. These are the numbers that decide:

• Core ML (fp16, iOS) — measured in the units that matter for this model — fraction of keypoints landing within 4 px of fp32: median 1.0000 over 8 real images, worst 1.0000.

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, 21, 512]	3.867e-06	1.000000
1	[1, 21, 512]	4.172e-06	1.000000

XNNPACK delegate coverage (fp32): 93.9% (306/326 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

Conversion

torch.export -> to_edge_transform_and_lower(partitioner) -> .pte (conversion scripts: executorch-models)