ONNX Model Support

[ebowwa]

Feature Request: ONNX Model Support

Enable CoreMLPlayer to load and run ONNX (Open Neural Network Exchange) models in addition to CoreML models.

Motivation

Bridge the training → deployment pipeline:

• Train models on GPU using edge-training platform
• Export to ONNX format
• Load directly in CoreMLPlayer for testing/validation
• Deploy to iOS/Mac after validation

Benefits:

• Cross-platform model compatibility
• Access to wider model ecosystem (HuggingFace, PyTorch Hub)
• Eliminates manual conversion steps during development
• Aligns with existing edge-training export capabilities (NCNN, ONNX, CoreML)

Proposed Implementation

Option A: Runtime Conversion (Recommended)

// On import, convert ONNX → CoreML
func loadONNXModel(url: URL) -> VNCoreMLModel {
    let coremlModel = convertONNXtoCoreML(url)
    return try VNCoreMLModel(for: coremlModel)
}

Pros:

• Native ANE/GPU acceleration
• No runtime dependencies
• Consistent performance with pure CoreML

Cons:

• Initial conversion delay
• Some ONNX ops may not convert

Option B: Direct ONNX Runtime

// Run ONNX directly using ort crate
func loadONNXModel(url: URL) -> ONNXSession {
    return try ONNXSession(modelPath: url.path)
}

Pros:

• No conversion overhead
• Support for latest ONNX ops

Cons:

• No ANE acceleration
• Additional dependency
• Slower inference

Acceptance Criteria

• Load .onnx files via CoreMLModelView
• Auto-convert to CoreML format on import (Option A)
• Run inference with same DetectionView overlay
• Display model metadata (ONNX opset version, inputs/outputs)
• Handle conversion errors gracefully with user feedback

Related

• feat: Integration with edge-training platform (GPU training → CoreML export) #5: edge-training platform integration
• feat: Real-time object detection with Meta Glasses (relay) #4: Meta Glasses real-time detection relay
• edge-training export formats: NCNN, TFLite, CoreML, ONNX

References

• coremltools ONNX conversion
• ONNX Format
• ONNX Runtime for macOS