pytorch faster rcnn 权重文件 导出 ONNX 格式

import torch
from torchvision import models


def mytranslate_pth_onnx(save_weight_pth, num=1, channel=3, height=512, width=672):
    """
    Torch Version 1.8.2+cu101
    :param save_weight_pth:  权重文件
    :param num:  1
    :param channel:  3
    :param height:  512
    :param width:  672
    :return:
    """

    print('Torch Version', torch.__version__)
    print('格式转换中...')

    # 模拟数据 N C H W
    example = torch.rand(num, channel, height, width)
    # 网络模型
    model = models.detection.fasterrcnn_resnet50_fpn(pretrained=True)
    checkpoint = torch.load(save_weight_pth, map_location=lambda storage, loc: storage.cuda(0))
    model.load_state_dict(checkpoint)
    # 导出ONNX
    model.eval()
    torch.onnx.export(model, example, save_weight_pth + ".onnx", verbose=True, opset_version=11)
    print("格式转换完成!")


if __name__ == "__main__":
    # 权重文件
    save_weight_pth = r'C:\Users\admin\.cache\torch\hub\checkpoints\fasterrcnn_resnet50_fpn_coco-258fb6c6.pth'
    num = 1
    channel = 3
    height = 512
    width = 672
    mytranslate_pth_onnx(save_weight_pth, num=1, channel=3, height=512, width=672)