v1.0.0

First tagged release. Self-contained ONNX export/inference pipeline running on current PyTorch 2.x, plus a Transformers.js example on the latest majors.

Highlights

• Self-contained ONNX pipeline — the MobileNetV2 backbone from upstream ZHKKKe/MODNet (Apache-2.0) is vendored under src/, so export/inference no longer require the upstream repo checked out.
• PyTorch 2.x support — runs on torch>=2.12 with CPU fallback; adds onnxscript (required by the 2.x exporter) and pins the legacy exporter (dynamo=False) so weights are embedded correctly (~25 MB) and dynamic_axes is honored.
• Dropped SwanLab — the export/inference scripts use plain print; no mandatory account or network dependency.
• Transformers.js v4 — examples/ upgraded to @huggingface/transformers 4 and typescript 6 (example code unchanged, type-checks clean).

Python / ONNX tooling (onnx_tools/)

• CPU fallback and DataParallel-prefix stripping in export_onnx.py; --opset-version flag (default 17).
• torch.load hardened with map_location + weights_only=True.
• inference_onnx.py: explicit CPUExecutionProvider, clearer tensor layout, required path args.
• Renamed onnx/ -> onnx_tools/ to stop the local package from shadowing the installed onnx library.

Dependencies

• Python: torch>=2.12, torchvision>=0.27, onnxruntime>=1.23.2, numpy>=2.2, latest onnx/opencv/pillow, add onnxscript, drop swanlab.
• JS/TS: @huggingface/transformers 3 -> 4, typescript 5 -> 6, tsx updated.

Docs

• Rewritten onnx_tools/README.md (virtualenv + install steps, argument tables, "Notes for PyTorch 2.x export").
• Root README.md entry point with a contents table and upstream credit; src/README.md attribution.

Verification

End-to-end export + inference on CPU: exported model ~24.76 MB (full weights), inference produces a valid alpha matte.