pytorch模型部署 pth转onnx的方法
作者:aoyou19
Pytorch转ONNX的意义
一般来说转ONNX只是一个手段,在之后得到ONNX模型后还需要再将它做转换,比如转换到TensorRT上完成部署,或者有的人多加一步,从ONNX先转换到caffe,再从caffe到tensorRT。Pytorch自带的torch.onnx.export转换得到的ONNX,ONNXRuntime需要的ONNX,TensorRT需要的ONNX都是不同的。
将pytorch训练保存的pth文件转为onnx文件,为后续模型部署做准备。
一、分类模型
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)
运行结果:
classify_model.pth model loaded
Exporting .pth model to onnx model has been successful!
Onnx model save as classify_model.onnxProcess finished with exit code 0
二、分割模型
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)
运行结果:
segment_model.pth model loaded
Exporting .pth model to onnx model has been successful!
Onnx model save as segment_model.onnxProcess finished with exit code 0
三、目标检测模型
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)
运行结果:
yolo.pth model loaded
Starting export with onnx 1.11.0.
Onnx model save as yolo.onnxProcess finished with exit code 0
参考链接:
1.yolo
2.模型部署翻车记:pytorch转onnx踩坑实录
到此这篇关于pytorch模型部署 pth转onnx的文章就介绍到这了,更多相关pytorch模型部署内容请搜索脚本之家以前的文章或继续浏览下面的相关文章希望大家以后多多支持脚本之家!