pytorch模型部署 pth转onnx的方法

import torch
import os
import timm
import argparse
from utils_net import Resnet
parser = argparse.ArgumentParser()
parser.add_argument("--pth_path", default='classify_model.pth')
parser.add_argument("--save_onnx_path", default='classify_model.onnx')
parser.add_argument("--input_width", default=416)
parser.add_argument("--input_height", default=416)
parser.add_argument("--input_channel", default=1)
parser.add_argument("--num_classes", default=6)
args = parser.parse_args()
def pth_to_onnx(pth_path, onnx_path, in_hig, in_wid, in_chal, num_cls):
    if not onnx_path.endswith('.onnx'):
        print('Warning! The onnx model name is not correct,\
              please give a name that ends with \'.onnx\'!')
        return 0
    model = Resnet(num_classes=num_cls)
    model.load_state_dict(torch.load(pth_path))
    model.eval()
    print(f'{pth_path} model loaded')
    input_names = ['input']
    output_names = ['output']
    im = torch.rand(1, in_chal, in_hig, in_wid)
    torch.onnx.export(model, im, onnx_path,
                      verbose=False,
                      input_names=input_names,
                      output_names=output_names)
    print("Exporting .pth model to onnx model has been successful!")
    print(f"Onnx model save as {onnx_path}")
if __name__ == '__main__':
    pth_to_onnx(pth_path=args.pth_path,
                onnx_path=args.save_onnx_path,
                in_hig=args.input_height,
                in_wid=args.input_width,
                in_chal=args.input_channel,
                num_cls=args.num_classes)

import torch
import os
import argparse
from utils_net import seg_net
parser = argparse.ArgumentParser()
parser.add_argument("--pth_path", default='segment_model.pth')
parser.add_argument("--save_onnx_path", default='segment_model.onnx')
parser.add_argument("--input_width", default=416)
parser.add_argument("--input_height", default=416)
parser.add_argument("--input_channel", default=1)
parser.add_argument("--num_classes", default=4)
args = parser.parse_args()
def pth_to_onnx(pth_path, onnx_path, in_hig, in_wid, in_channel, num_cls):
    if not onnx_path.endswith('.onnx'):
        print('Warning! The onnx model name is not correct,\
              please give a name that ends with \'.onnx\'!')
        return 0
    model = seg_net(in_channel=in_channel, num_cls=num_cls)
    model.load_state_dict(torch.load(pth_path))
    model.eval()
    print(f'{pth_path} model loaded')
    input_names = ['input']
    output_names = ['output']
    im = torch.rand(1, in_channel, in_hig, in_wid)
    torch.onnx.export(model, im, onnx_path,
                      verbose=False,
                      input_names=input_names,
                      output_names=output_names,
                      opset_version=11)
    print("Exporting .pth model to onnx model has been successful!")
    print(f"Onnx model save as {onnx_path}")
if __name__ == '__main__':
    pth_to_onnx(pth_path=args.pth_path,
                onnx_path=args.save_onnx_path,
                in_hig=args.input_height,
                in_wid=args.input_width,
                in_channel=args.input_channel,
                num_cls=args.num_classes)

在这里插入代码片
import torch
import onnx
import argparse
from utils_net import YoloBody
parser = argparse.ArgumentParser()
parser.add_argument("--pth_path", default='yolo.pth')
parser.add_argument("--save_onnx_path", default='yolo.onnx')
parser.add_argument("--input_width", default=416)
parser.add_argument("--input_height", default=416)
parser.add_argument("--num_classes", default=2)
parser.add_argument("--anchors_mask", default=[[6, 7, 8], [3, 4, 5], [0, 1, 2]])
args = parser.parse_args()
def pth_to_onnx(pth_path: str, save_onnx_path: str, num_cls: int,
                in_hig: int, in_wid: int, anchor_mask: list,
                opset_version: int = 12, simplify: bool = False):
    """
    :param pth_path: pth文件文件
    :param save_onnx_path: 准备保存的onnx路径
    :param num_cls: 检测目标类别数
    :param in_hig: 网络输入高度
    :param in_wid: 网络输入宽度
    :param anchor_mask: anchor宽高索引
    :param opset_version: onnx算子集版本
    :param simplify: 是否对模型进行简化
    :return:保存onnx到指定路径
    """
    # Build model, load weights
    net = YoloBody(anchors_mask=anchor_mask,
                   num_classes=num_cls)
    # device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    # net.load_state_dict(torch.load(pth_path, map_location=device))
    net.load_state_dict(torch.load(pth_path))
    # print(next(net.parameters()).device)
    net = net.eval()
    print(f'{pth_path} model loaded')
    im = torch.zeros(1, 3, in_hig, in_wid).to('cpu')
    input_layer_names = ['images']
    output_layer_names = ['output']
    # Export the model
    print(f'Starting export with onnx {onnx.__version__}.')
    torch.onnx.export(net,
                      im,
                      f=save_onnx_path,
                      verbose=False,
                      opset_version=opset_version,
                      training=torch.onnx.TrainingMode.EVAL,
                      do_constant_folding=True,
                      input_names=input_layer_names,
                      output_names=output_layer_names,
                      dynamic_axes=None)
    # Checks
    model_onnx = onnx.load(save_onnx_path)  # load onnx model
    onnx.checker.check_model(model_onnx)  # check onnx model
    # Simplify onnx
    if simplify:
        import onnxsim
        print(f'Simplifying with onnx-simplifier {onnxsim.__version__}.')
        model_onnx, check = onnxsim.simplify(
            model_onnx,
            dynamic_input_shape=False,
            input_shapes=None)
        assert check, 'assert check failed'
        onnx.save(model_onnx, save_onnx_path)
    print('Onnx model save as {}'.format(save_onnx_path))
if __name__ == '__main__':
    pth_to_onnx(pth_path=args.pth_path,
                save_onnx_path=args.save_onnx_path,
                num_cls=args.num_classes,
                in_hig=args.input_height,
                in_wid=args.input_width,
                anchor_mask=args.anchors_mask)