onnx模型转换opset版本和固定动态输入尺寸

背景：之前我想把onnx模型从opset12变成opset12，太慌乱就没找着，最近找到了官网上有示例的，大爱onnx官网，分享给有需求没找着的小伙伴们。

1. onnx模型转换opset版本

官网示例：

1. import onnx

2. from onnx import version_converter, helper

4. # Preprocessing: load the model to be converted.

5. model_path = "path/to/the/model.onnx"

6. original_model = onnx.load(model_path)

8. print(f"The model before conversion:\n{original_model}")

10. # A full list of supported adapters can be found here:

11. # https://github.com/onnx/onnx/blob/main/onnx/version_converter.py#L21

12. # Apply the version conversion on the original model

13. converted_model = version_converter.convert_version(original_model, <int target_version>)

15. print(f"The model after conversion:\n{converted_model}")

其github地址如下：

onnx/docs/PythonAPIOverview.md at main · onnx/onnx (github.com)​编辑https://github.com/onnx/onnx/blob/main/docs/PythonAPIOverview.md#converting-version-of-an-onnx-model-within-default-domain-aionnx其小伙伴拉到gitee上的地址如下（以防有的小伙伴github打不开）：

docs/PythonAPIOverview.md · meiqicheng/github-onnx-onnx - Gitee.com​编辑https://gitee.com/meiqicheng1216/onnx/blob/master/docs/PythonAPIOverview.md#converting-version-of-an-onnx-model-within-default-domain-aionnx最后附上完整代码：

1. import onnx

2. from onnx import version_converter, helper

4. # A full list of supported adapters can be found here:

5. # https://github.com/onnx/onnx/blob/main/onnx/version_converter.py#L21

6. # Apply the version conversion on the original model

8. # Preprocessing: load the model to be converted.

9. model_path = r"./demo.onnx"

10. original_model = onnx.load(model_path)

11. print(f"The model before conversion:\n{original_model}")

14. converted_model = version_converter.convert_version(original_model, 11)

15. print(f"The model after conversion:\n{converted_model}")

17. save_model = model_path[:-5] + "_opset11.onnx"

18. onnx.save(converted_model, save_model)

2. onnx模型转固定动态输入尺寸

1. def change_dynamic_input_shape(model_path, shape_list: list):

2. """

3. 将动态输入的尺寸变成固定尺寸

4. Args:

5. model_path: onnx model path

6. shape_list: [1, 3, ...]

7. Returns:

9. """

10. import os

11. import onnx

12. model_path = os.path.abspath(model_path)

13. output_path = model_path[:-5] + "_fixed.onnx"

14. model = onnx.load(model_path)

15. # print(onnx.helper.printable_graph(model.graph))

16. inputs = model.graph.input # inputs是一个列表，可以操作多输入~

17. # look_input = inputs[0].type.tensor_type.shape.dim

18. # print(look_input)

19. # print(type(look_input))

20. # inputs[0].type.tensor_type.shape.dim[0].dim_value = 1

21. for idx, i_e in enumerate(shape_list):

22. inputs[0].type.tensor_type.shape.dim[idx].dim_value = i_e

23. # print(onnx.helper.printable_graph(model.graph))

24. onnx.save(model, output_path)

27. if __name__ == "__main__":

28. model_path = "./demo.onnx"

29. shape_list = [1]

30. change_dynamic_input_shape(model_path, shape_list)

---

---

---

onnx模型转换opset版本和固定动态输入尺寸_onnx opset-CSDN博客