classification-models-pytorch/export_onnx.py

import torch
import torchvision

import onnx


def get_args_parser(add_help=True):
    import argparse

    parser = argparse.ArgumentParser(description="model weight transport to onnx", add_help=add_help)

    parser.add_argument("--model", default="resnet18", type=str, help="model name")
    parser.add_argument(
        "--num_classes", default=10, type=int, help="number of classes"
    )
    parser.add_argument(
        "--input_size", default=224, type=int, help="input size"
    )
    parser.add_argument("--weights", default=None, type=str, help="the weights enum name to load")
    parser.add_argument("--save_path", default=None, type=str, help="onnx file save path")
    return parser


def export_onnx(args):
    # 加载模型
    model = torchvision.models.get_model(args.model, weights=None, num_classes=args.num_classes)
    model.eval()  # 切换到评估模式

    # 加载权重
    checkpoint = torch.load(args.weights, map_location='cpu')
    model.load_state_dict(checkpoint["model"])

    # 定义一个随机输入张量，尺寸应符合模型输入要求
    # 通常，ImageNet预训练模型的输入尺寸是 (batch_size, 3, 224, 224)
    dummy_input = torch.randn(1, 3, args.input_size, args.input_size)

    # 导出模型到 ONNX 格式
    with torch.no_grad():
        torch.onnx.export(
            model,
            dummy_input,
            args.save_path,
            export_params=True,
            opset_version=11,
            do_constant_folding=False,  # 确保这个参数一定是False，如果不为False，导出的onnx权重与原始权重数值不一致，导致白盒水印提取失败
            input_names=["input"],
            output_names=["output"],
            dynamic_axes={"input": {0: "batch_size"}, "output": {0: "batch_size"}}
        )

    print("模型成功导出为 ONNX 格式！")

    # Checks
    model_onnx = onnx.load(args.save_path)  # load onnx model
    onnx.checker.check_model(model_onnx)  # check onnx model


if __name__ == '__main__':
    args = get_args_parser().parse_args()
    export_onnx(args)