| --- |
| license: apache-2.0 |
| tags: |
| - executorch |
| - xnnpack |
| - pte |
| - on-device |
| - keypoint-detection |
| - pose-estimation |
| --- |
| # rtmpose_m_face — ExecuTorch |
|
|
| - **Source**: open-mmlab/mmpose RTMPose (rtmpose-m_simcc-face6_pt-in1k_120e-256x256-72a37400) |
| - **License**: Apache-2.0 |
| - **Input**: [[1, 3, 256, 256]] — RGB, ImageNet norm, 256x256 crop around a face (detect first, then crop and resize to this aspect) |
| - **Output**: SimCC pair: x [1,106,512] and y [1,106,512] — 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_face_xnnpack_fp32.pte` | 67.9 | 1.000000 | 10.4 | |
| | Core ML (fp16, iOS) | `rtmpose_m_face_coreml_all.pte` | 34.4 | 0.999897 | 3.3 | |
| |
| |
| 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: 94.4 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 0.9811 over 10 real images, worst 0.9151. |
| |
| ## 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, 106, 512] | 1.147e-05 | 1.000000 | |
| | 1 | [1, 106, 512] | 9.179e-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](https://github.com/john-rocky/executorch-models)) |
|
|
| **Notes**: Top-down: crop one face first. In practice a portrait framed on the head works directly, which is how the card check exercises it. |
|
|
| <!-- funnel:v1 --> |
|
|
| --- |
|
|
| **More models in this format:** [ExecuTorch Model Zoo](https://huggingface.co/collections/mlboydaisuke/executorch-model-zoo-6a7ff328390b63075ffeae5e) — 31 models, each with the recipe that produced it. |
|
|
| **Want a different model on-device?** [Open a request](https://github.com/john-rocky/on-device-requests) — free, open weights only; the export and its measured numbers get published publicly. |
|
|
| <!-- /funnel:v1 --> |
|
|
