Upload README.md with huggingface_hub

README.md

@@ -21,14 +21,25 @@ All variants take and return fp32 tensors — swap the `.pte` file, keep your ap
 
 | build | file | size (MB) | parity vs fp32 eager (worst corr) | Mac median (ms)* |
 |-----------|------|-----------|------------------------------------|------------------|
-| fp32 | `rtmpose_m_animal_xnnpack_fp32.pte` | 54.5 | 1.000000 | 9.4 |
+| fp32 | `rtmpose_m_animal_xnnpack_fp32.pte` | 54.5 | 1.000000 | 9.3 |
+| Core ML (fp16, iOS) | `rtmpose_m_animal_coreml_all.pte` | 27.7 | 0.998051 | 2.8 |
+
+
+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: 91.4 ms).
+only, not a device number (torch eager fp32 on the same machine: 93.4 ms).
+
+### Checked in the task's own units
 
-### Builds that did not earn a slot
+Correlation is a first filter. These are the numbers that decide:
 
-- **Core ML (fp16, iOS) is not shipped**: measured in the units that matter for this model — fraction of keypoints landing within 4 px of fp32: median 0.9412 over 10 real images, worst 0.7647.
+- **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)
 
@@ -37,8 +48,8 @@ the correlation over all elements of each output tensor.
 
 | output | shape | max_abs_diff | corr |
 |--------|-------|--------------|------|
-| 0 | [1, 17, 512] | 2.783e-06 | 1.000000 |
-| 1 | [1, 17, 512] | 7.629e-06 | 1.000000 |
+| 0 | [1, 17, 512] | 7.242e-06 | 1.000000 |
+| 1 | [1, 17, 512] | 4.828e-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
 
@@ -46,15 +57,3 @@ XNNPACK delegate coverage (fp32): 93.9% (306/326 ops); ops left on the portable
 
 torch.export -> to_edge_transform_and_lower(partitioner) -> .pte
 (conversion scripts: [executorch-models](https://github.com/john-rocky/executorch-models))
-
-**Notes**: Top-down again: crop one animal first. AP-10K covers 54 mammal species; a general object detector's animal classes make a workable front end.
-
-<!-- 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 -->