Skip to content

refactor: standardize on onnx-community HuggingFace format#167

Closed
solderzzc wants to merge 2 commits intodevelopfrom
feature/onnx-coreml-inference
Closed

refactor: standardize on onnx-community HuggingFace format#167
solderzzc wants to merge 2 commits intodevelopfrom
feature/onnx-coreml-inference

Conversation

@solderzzc
Copy link
Member

  • Replace ultralytics-exported yolo26n.onnx with onnx-community version
  • Simplify _OnnxCoreMLModel to single format: logits [1,300,80] + pred_boxes [1,300,4]
  • Remove dual-format detection and ultralytics output parsing
  • All YOLO26 sizes (n/s/m/l) now use the same onnx-community format
  • Verified: 15-25ms/frame on M5 Pro with CoreML EP, correct detections

- Ship yolo26n.onnx (9.4MB) from onnx-community for zero-config default
- On-demand download for s/m/l from onnx-community HuggingFace repos
- _OnnxCoreMLModel wrapper: logits+pred_boxes → sigmoid+cxcywh→xyxy
- Bypasses macOS 26.x MPSGraph MLIR crash (SIGABRT)
- 15-25ms/frame (~67 FPS) on Apple M5 Pro with CoreML EP
- Strip requirements_mps.txt: no torch/ultralytics (~120MB → ~17MB)
- Class names from yolo26n_names.json (no .pt dependency)
@solderzzc solderzzc force-pushed the feature/onnx-coreml-inference branch from 7a5d1be to 76ede38 Compare March 18, 2026 19:14
@solderzzc solderzzc closed this Mar 18, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant