scripts/export_onnx.py

"""
============================================================
Export Model to ONNX for Deployment
============================================================
Usage:
    python scripts/export_onnx.py --checkpoint results/checkpoints/resnet50_best.pth
============================================================
"""

import os
import sys
import yaml
import argparse
import torch
import torch.onnx

sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from models.classifier import build_model


def export_to_onnx(checkpoint_path, config_path="configs/config.yaml", output_path=None):
    with open(config_path) as f:
        config = yaml.safe_load(f)
    
    checkpoint = torch.load(checkpoint_path, map_location="cpu")
    model_name = checkpoint["model_name"]
    
    model = build_model(model_name, config)
    model.load_state_dict(checkpoint["state_dict"])
    model.eval()
    
    if output_path is None:
        output_path = checkpoint_path.replace(".pth", ".onnx")
    
    dummy_input = torch.randn(1, 3, 224, 224)
    
    torch.onnx.export(
        model, dummy_input, output_path,
        export_params=True,
        opset_version=14,
        do_constant_folding=True,
        input_names=["input"],
        output_names=["output"],
        dynamic_axes={
            "input": {0: "batch_size"},
            "output": {0: "batch_size"},
        },
    )
    
    # Verify
    import onnx
    onnx_model = onnx.load(output_path)
    onnx.checker.check_model(onnx_model)
    
    size_mb = os.path.getsize(output_path) / (1024 ** 2)
    print(f"Exported: {output_path} ({size_mb:.1f} MB)")
    
    # Test with ONNX Runtime
    try:
        import onnxruntime as ort
        session = ort.InferenceSession(output_path)
        result = session.run(None, {"input": dummy_input.numpy()})
        print(f"ONNX Runtime test: output shape = {result[0].shape}")
    except ImportError:
        print("Install onnxruntime to verify: pip install onnxruntime")


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--checkpoint", type=str, required=True)
    parser.add_argument("--config", type=str, default="configs/config.yaml")
    parser.add_argument("--output", type=str, default=None)
    args = parser.parse_args()
    
    export_to_onnx(args.checkpoint, args.config, args.output)