User Guide: Model Export & Edge Optimization

CoreCV features a dedicated, multi-stage model export engine managed by CoreExporter, TargetRewriter, and MetaProber. It ensures smooth compilation and transition from PyTorch models into edge-optimized ONNX and ExecuTorch formats.

---

1. Complete Parameter Reference

CoreModel.export() & CoreExporter Parameters

Parameter	Type Hint	Default	Description
format	str	"onnx"	Target export format. One of "onnx", "executorch", or "both".
target_hardware	str	"server"	Hardware profile ("edge" or "server"). "edge" applies activation rewrites (GELU→ReLU, SiLU→Hardswish) and NCHW layout optimizations; "server" skips rewrites.
opset	int	17	ONNX opset version. Valid options: 17 or 18.
optimize	bool	True	Applies TargetRewriter graph optimization passes and XNNPACK delegates (for ExecuTorch).
output_path	str \| None	None	Explicit output file path. For "both", used as a prefix (e.g. "model" produces "model.onnx" and "model.pte"). Auto-generated if None.
input_shape	tuple[int, ...]	(1, 3, H, W)	Dummy input tensor dimensions (B, C, H, W) used for graph tracing and shape auditing.
dynamic_axes	dict \| None	None	ONNX dynamic axes mapping, e.g. {"input": {0: "batch", 2: "height", 3: "width"}}.
weights	str \| Path \| None	None	Optional path to a .pt/.pth checkpoint file loaded into the model before export.

---

2. Multi-Stage Edge Export Pipeline

  CoreModel (PyTorch nn.Module)
               │
               ▼
   1. TargetRewriter (Edge Hardware Transformations)
      - Activation Rewrites: GELU -> ReLU, SiLU -> Hardswish
      - LayerNorm Collapse & Constant Folding
               │
               ▼
   2. MetaProber (Zero-VRAM Shape & Op Verification)
      - Dynamic operation audit on 'meta' device tensors
               │
               ▼
   3. Serialized Model Export
      ├── ONNX Export (.onnx) via torch.onnx.export
      └── ExecuTorch Export (.pte) via torch.export

---

3. Practical Examples: Detection vs. Segmentation Export

Object Detection Export (ONNX & ExecuTorch)Semantic Segmentation Export (Dual Format)

from corecv.api import CoreModel

# 1. Initialize Object Detector Facade
model = CoreModel(
    model="configs/yolo_detection.yaml",
    task="detection",
    input_size=(640, 640),
)

# 2. Export to Edge-Optimized ONNX format with Dynamic Axes
onnx_result = model.export(
    format="onnx",
    target_hardware="edge",  # Applies GELU->ReLU & SiLU->Hardswish rewrites
    opset=18,
    dynamic_axes={"input": {0: "batch_size"}},
    output_path="exports/yolo_detection_edge.onnx",
    weights="runs/train/exp1/best.pt",
)
print(f"ONNX Model saved to: {onnx_result['onnx']}")

# 3. Export to ExecuTorch (.pte) format
pte_result = model.export(
    format="executorch",
    target_hardware="edge",
    output_path="exports/yolo_detection_edge.pte",
)
print(f"ExecuTorch Model saved to: {pte_result['executorch']}")

from corecv.api import CoreModel

# 1. Initialize Segmentation Model Facade
model = CoreModel(
    model="configs/deeplabv3_segm.yaml",
    task="segmentation",
    input_size=(512, 512),
)

# 2. Export Both Formats Simultaneously
export_paths = model.export(
    format="both",
    target_hardware="edge",
    output_path="exports/segmentation_model",  # Produces segmentation_model.onnx & segmentation_model.pte
    weights="checkpoints/segm_best.pt",
)

print(f"ONNX path: {export_paths['onnx']}")
print(f"ExecuTorch path: {export_paths['executorch']}")