export_onnx.py

"""Export TwinLiteNet8 to ONNX for cross-platform deployment.

Usage:
    python export_onnx.py --ckpt run_8class/twinlite8_best.pt --out twinlite8.onnx
    python export_onnx.py --ckpt run_8class/twinlite8_best.pt --out twinlite8_dynamic.onnx --dynamic
"""
import argparse, sys, os
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from pathlib import Path
import numpy as np, torch
from model.TwinLite_8class import TwinLiteNet8


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--ckpt", required=True)
    ap.add_argument("--out", required=True)
    ap.add_argument("--height", type=int, default=360)
    ap.add_argument("--width",  type=int, default=640)
    ap.add_argument("--dynamic", action="store_true",
                    help="Allow dynamic batch + spatial dims (slightly slower at runtime)")
    ap.add_argument("--opset", type=int, default=17)
    args = ap.parse_args()

    print(f"loading ckpt: {args.ckpt}")
    model = TwinLiteNet8(num_classes=8).eval()
    ckpt = torch.load(args.ckpt, map_location="cpu", weights_only=False)
    model.load_state_dict(ckpt["model"])
    print(f"  epoch {ckpt['epoch']}  tree IoU {ckpt.get('tree_iou_old','?')}")

    dummy = torch.randn(1, 3, args.height, args.width)

    if args.dynamic:
        dyn = {"input": {0: "batch", 2: "height", 3: "width"},
               "output": {0: "batch", 2: "height", 3: "width"}}
    else:
        dyn = None

    print(f"exporting to ONNX (opset {args.opset}) ...")
    torch.onnx.export(
        model, dummy, args.out,
        input_names=["input"], output_names=["output"],
        dynamic_axes=dyn,
        opset_version=args.opset,
        do_constant_folding=True,
    )

    sz = os.path.getsize(args.out) / 1e6
    print(f"  saved: {args.out}  ({sz:.2f} MB)")

    # Validate ONNX numerical parity vs PyTorch
    try:
        import onnxruntime as ort
        sess = ort.InferenceSession(args.out, providers=["CPUExecutionProvider"])
        with torch.no_grad():
            torch_out = model(dummy).numpy()
        onnx_out = sess.run(None, {"input": dummy.numpy()})[0]
        diff = np.abs(torch_out - onnx_out)
        argmax_match = (torch_out.argmax(1) == onnx_out.argmax(1)).mean()
        print(f"  parity: max_abs_diff={diff.max():.6f}  mean={diff.mean():.6f}")
        print(f"  argmax agreement: {100*argmax_match:.4f}%  (must be ~100% for safe deploy)")
        assert argmax_match > 0.999, "argmax disagreement > 0.1% — investigate"
        print("  PARITY OK")
    except ImportError:
        print("  (skip parity check — onnxruntime not installed; pip install onnxruntime)")


if __name__ == "__main__":
    main()