7. Export & Deployment

Export Modes (python -m src.export)

Mode	Output	Notes
onnx	<backbone>_fp32.onnx	opset 13, static or dynamic batch
int8	<backbone>_int8.pt + _int8_traced.pt	PTQ with 32-batch calibration
float16	<backbone>_float16.pt	TorchScript traced
portable	portable/<backbone>_<tag>/ folder	Self-contained: model + labels + info

Portable Export Structure

outputs/export/portable/<backbone>_phase2_best/
├── model.pt                 # torch checkpoint {state_dict: ...}
├── cattle_classes.json      # {"amritmahal": 0, "ayrshire": 1, ...}
├── buffalo_classes.json     # {"alambadi": 0, "banni": 1, ...}
└── model_info.json          # {backbone, image_size, usage, exported_at}

---