feat: Integration with edge-training platform (GPU training → CoreML export)

[ebowwa]

Feature: Integration with edge-training Platform

Connect CoreMLPlayer with the edge-training platform to create a complete training-to-deployment pipeline.

The Vision

edge-training (GPU) → Export → CoreMLPlayer → Deploy on Mac/Glasses
     ↓ Training
     ↓ Optimization (MobileNet, pruning, etc.)
     ↓ Export to CoreML
     ↓ Open in CoreMLPlayer
     ↓ Run inference on live video/glasses

What edge-training Provides

Training Infrastructure:

• Multi-GPU YOLO/RT-DETR training
• Cloud GPU provider integrations (RunPod, Kaggle)
• Automated dataset prep and structure detection

Optimization:

• MobileNet convolutions (~8x speedup)
• Attention head pruning
• Knowledge distillation
• OFA (Once-for-All) networks

Export Formats:

• CoreML (for Mac/glasses)
• NCNN, TFLite, ONNX (for other edge devices)

Integration Opportunity

Option 1: Direct Export

• Add "Export to CoreML" button in edge-training
• Uses existing export_service.py
• Auto-generates .mlmodelc files

Option 2: Model Registry

• edge-training exports to model registry
• CoreMLPlayer browses and downloads models
• Version tracking and metadata

Option 3: Live Feedback Loop

• CoreMLPlayer captures inference results
• Send back to edge-training for retraining
• Continuous improvement pipeline

Technical Notes

edge-training already has:

• ✅ service/export_service.py - export to CoreML/NCNN/TFLite
• ✅ Model training with Ultralytics YOLO
• ✅ Edge optimization modules
• ✅ Multi-format export support

CoreMLPlayer has:

• ✅ .mlmodel/.mlpackage loading
• ✅ Real-time video inference
• ✅ Detection visualization
• ✅ Live video feed support (could add glasses feed)

Use Cases

1. Train on GPU → Optimize → Deploy on Mac (CoreMLPlayer)
2. Glasses capture data → Train in cloud → Push back to glasses
3. Custom object → Train YOLO → Test in CoreMLPlayer with live glasses

Repositories

• edge-training: https://github.com/ebowwa/edge-training
• CoreMLPlayer: https://github.com/ebowwa/CoreMLPlayer

🤖 Generated with Claude Code