Align model card with the established ExecuTorch model-card format

README.md

@@ -1,102 +1,62 @@
 ---
 license: apache-2.0
 pipeline_tag: image-segmentation
-tags:
-  - executorch
-  - react-native
-  - background-removal
-  - salient-object-detection
+library_name: executorch
 ---
 
-# U²-Netp for React Native ExecuTorch
+# u2netp
 
-The small U²-Net salient object model, adapted for the existing
-`createSemanticSegmenter` / `useSemanticSegmenter` pipeline in
-[React Native ExecuTorch](https://github.com/software-mansion/react-native-executorch).
-The source model and pretrained checkpoint are from
-[xuebinqin/U-2-Net](https://github.com/xuebinqin/U-2-Net), under Apache-2.0.
+This repository hosts the **U²-Netp** models exported for the
+[React Native ExecuTorch](https://www.npmjs.com/package/react-native-executorch)
+library as ExecuTorch `.pte` programs, ready to run on device.
 
-This model predicts salient foreground objects, including products and animals,
-as well as people. It produces one soft foreground mask; it does not classify
-object categories or separate individual instances. Fine structures and
-low-contrast boundaries can be missed.
+U²-Netp predicts a soft mask for salient foreground objects, including animals
+and products as well as people. It does not classify object categories or
+separate individual instances.
 
-## Files
+Upstream model: [U²-Net](https://github.com/xuebinqin/U-2-Net),
+revision `ac7e1c817ecab7c7dff5ce6b1abba61cd213ff29`.
 
-| Backend | File | Size | Compute precision |
-| --- | --- | --- | --- |
-| XNNPACK | `xnnpack/u2netp_xnnpack_fp32.pte` | 4.6 MB | FP32 |
-| CoreML | `coreml/u2netp_coreml_fp16.pte` | 3.3 MB | FP16 |
-| CoreML | `coreml/u2netp_coreml_fp32.pte` | 5.5 MB | FP32 |
+## Variants
 
-Root `config.json` declares `modelName: u2netp`. Each backend has its own
-`config.json` with the exported tensor signatures. `sha256.json` records the
-artifact checksums. Models were exported with torch 2.14.0, ExecuTorch 1.4.1 and
-coremltools 9.0. CoreML targets iOS 18 / macOS 15 or newer.
+| Path                              | Backend | Precision |
+| --------------------------------- | ------- | --------- |
+| `coreml/u2netp_coreml_fp16.pte`   | coreml  | fp16      |
+| `coreml/u2netp_coreml_fp32.pte`   | coreml  | fp32      |
+| `xnnpack/u2netp_xnnpack_fp32.pte` | xnnpack | fp32      |
 
-## Tensor contract and processing
+The exports take RGB float32 `[1,3,320,320]` input in `[0,1]` and return one
+float32 `[1,1,320,320]` foreground logit map. Image-maximum scaling and ImageNet
+normalization are included in the graph. RNE applies sigmoid and returns an
+opaque RGBA grayscale mask; use its foreground intensity as an alpha mask when
+compositing the original image. Fine structures can be missed. FP16 favors
+speed and may shift mask boundaries; FP32 preserves the reference more closely.
 
-`forward(float32[1,3,320,320]) -> float32[1,1,320,320]`
+## Repository structure
 
-Input is RGB, stretched to 320×320, in NCHW layout with values in `[0,1]`.
-The exported graph divides by the image maximum (clamped to `1e-6` for an
-all-black input), then applies ImageNet mean/std normalization. Do not apply
-ImageNet normalization a second time on the client.
-
-Output is the fused foreground **logit** map. The seven upstream probability
-outputs are replaced by the fused logit before sigmoid, retaining the pretrained
-parameters. The client applies sigmoid once and resizes the mask to the source
-image dimensions. The upstream demo's per-image output min/max contrast
-stretch is omitted, so intensities remain foreground probabilities.
-
-In RNE, `models.semanticSegmentation.U2NETP` configures RGB scaling by `1/255`,
-linear input/output interpolation, and labels `background` / `foreground`.
-The semantic segmenter returns an RGBA grayscale mask: RGB contains foreground
-intensity from 0 to 255, alpha is opaque, and `colormap` is undefined. Use an RGB
-channel as the alpha mask when compositing the original image; the returned
-buffer itself is not an image cutout.
-
-## Validation and performance
-
-Host checks used an Apple M1 Pro running macOS 26.5.2. These are warm host
-measurements, not phone benchmarks, and exclude image preprocessing and mask
-rendering in React Native.
-
-XNNPACK runs through the ExecuTorch runtime. Its probability error against
-upstream eager PyTorch is below 0.00015 on the tested bicycle, boat and animal
-images. The graph contains one XNNPACK delegate plus a portable `amax` input
-reduction. Per-op timing was unavailable because the installed runtime wheel
-does not enable event tracing.
-
-The actual CoreML partitions were extracted from the saved `.pte` files and
-executed directly through CoreML, since the installed ExecuTorch wheel cannot
-execute its multifunction CoreML packages. Both CoreML exports contain one
-delegate with no portable operations. The profiler's compute-plan tool places
-369 of 370 dispatched FP16 operations on the Neural Engine; one cast runs on
-CPU.
-
-On the three test images, FP16 foreground-probability MAE is 0.00089–0.00453
-and thresholded-mask IoU against FP32 is 0.9873–0.9991. Individual edge pixels
-can differ by up to 0.312, so FP16 is a speed/quality tradeoff. CoreML FP32
-matches the eager thresholded masks on all three images, with maximum
-probability error below 0.00017. Warm direct-CoreML medians are approximately
-7.2 ms (FP16) and 16.7 ms (FP32).
-
-Black/white inputs were also checked for finite results. This is a conversion
-smoke test, not a dataset accuracy evaluation. Full CoreML execution through
-RNE and both backends' performance require validation on the target devices.
+```
+config.json                                  28 B
+coreml/config.json                         1.5 kB
+coreml/u2netp_coreml_fp16.pte                3.3 MB
+coreml/u2netp_coreml_fp32.pte                5.5 MB
+xnnpack/config.json                        888 B
+xnnpack/u2netp_xnnpack_fp32.pte              4.6 MB
+LICENSE                                  11.4 kB
+sha256.json                                314 B
+```
 
-## Reproduce the exports
+## Compatibility
 
-From the export-scripts repository:
+These files are published for the **ExecuTorch v1.4.1** runtime. ExecuTorch
+gives no forward compatibility guarantee, so an older runtime may fail to load
+them. CoreML requires iOS 18 / macOS 15 or newer; full RNE execution and
+performance still require validation on the target device.
 
-```sh
-uv run --project exporter_models/u2netp exporter export u2netp --backend xnnpack --output-dir exports/u2netp/xnnpack
-uv run --project exporter_models/u2netp exporter export u2netp --backend coreml --variant fp16 --output-dir exports/u2netp/coreml
-uv run --project exporter_models/u2netp exporter export u2netp --backend coreml --variant fp32 --output-dir exports/u2netp/coreml
-```
+To use them in React Native ExecuTorch, pass
+`models.semanticSegmentation.U2NETP.DEFAULT` to `useSemanticSegmenter` or
+`createSemanticSegmenter`. See the
+[documentation](https://docs.swmansion.com/react-native-executorch/docs/fundamentals/downloading-models).
 
-The model code is vendored from upstream commit
-`ac7e1c817ecab7c7dff5ce6b1abba61cd213ff29`. The checkpoint is downloaded from the
-author's linked Google Drive artifact and checked against SHA-256
-`e7567cde013fb64813973ce6e1ecc25a80c05c3ca7adbc5a54f3c3d90991b854`.
+To load these files in your own ExecuTorch runtime, read the
+[compatibility note](https://github.com/pytorch/executorch/blob/main/runtime/COMPATIBILITY.md)
+first.