tools for model exporting and converting #639

horcham · 2023-12-21T12:10:38Z

Thank you for your contribution to the MindOCR repo.
Before submitting this PR, please make sure:

You have read the Contributing Guidelines on pull requests
Your code builds clean without any errors or warnings
You are using approved terminology
You have added unit tests

Motivation

提供原生模型转换工具，用户终于可以无痛转模型了。支持以下功能
- 三种转换模式：export，convert，export+convert
- 支持动静态
- 支持下载
- 支持指定模型，或模型全量转换
- 支持convert后使用benchmark检验
- 输出转换指令与转换报告，方便检验与复现
- 保存转换日志，方便问题定位
使用例子：

# 全量模型，下载+转换，export+convert，动态
python export_convert_tool.py --task export_convert --model_name all --is_dynamic True --force True

# 指定模型，下载，export，静态
python export_convert_tool.py --task export --model_name crnn_vgg7 --is_dynamic False --force True

# 指定模型。指定mindir，convert，动态
python export_convert_tool.py --task convert--model_name crnn_vgg7  --input_file="crnn_vgg7.mindir" --is_dynamic True --force True

输出指定与报告

指定模型，运行后输出操作指令与报告

例1：运行export，得到如下指令和报告

 python export_convert_tool.py --model_name dbnetpp_resnet50 --task export --is_dynamic true --input_file dbnetpp_resnet50-068166c2.ckpt

Export Command: python export.py --model_name_or_config dbnetpp_resnet50 --save_dir output_folder/export_folder --local_ckpt_path dbnetpp_resnet50-068166c2.ckpt --is_dynamic_shape True --model_type det
Export Success: output_folder/export_folder/dbnetpp_resnet50.mindir
+----------------------+----------------+
| Export Dynamic Model | Dynamic Export |
+----------------------+----------------+
|   dbnetpp_resnet50   |    Success     |
+----------------------+----------------+

例2：运行convert，得到如下指令和报告

 python export_convert_tool.py --model_name dbnetpp_resnet50 --task export --is_dynamic true --input_file dbnetpp_resnet50-068166c2.ckpt

Convert Command: converter_lite --fmk=MINDIR --modelFile=output_folder/export_folder/dbnetpp_resnet50.mindir --outputFile=output_folder/convert_folder/dbnetpp_resnet50_lite_dynamic --optimize=ascend_oriented --configFile=output_folder/convert_folder/convert_dynamic_dbnetpp_resnet50.txt
Convert Success: output_folder/convert_folder/dbnetpp_resnet50_lite_dynamic.mindir
Benchmark Command: benchmark --modelFile=output_folder/convert_folder/dbnetpp_resnet50_lite_dynamic.mindir --device=Ascend --inputShapes=1,3,1152,2048 --loopCount=100 --warmUpLoopCount=10
Infer Dynamic Shape Success 
Benchmark Command: benchmark --modelFile=output_folder/convert_folder/dbnetpp_resnet50_lite_dynamic.mindir --device=Ascend --inputShapes=1,3,1120,2080 --loopCount=100 --warmUpLoopCount=10
Infer Dynamic Shape Success 

+--------------------------------------------+-----------------+---------------+---------------+
| Convert Dynamic Model From Exported Mindir | Dynamic Convert | Infer Shape 0 | Infer Shape 1 |
+--------------------------------------------+-----------------+---------------+---------------+
|              dbnetpp_resnet50              |     Success     | 1,3,1152,2048 | 1,3,1120,2080 |
+--------------------------------------------+-----------------+---------------+---------------+

例3：运行export_convert，全量动态转换，得到如下指令和报告

 python export_convert_tool.py --task export_convert --model_name all --is_dynamic True --force True

Exporting model:abinet, 1/24
output_folder/export_folder/abinet.mindir exists and it will be overwritten if exported successfully.
Export Command: python export.py --model_name_or_config abinet --save_dir output_folder/export_folder --is_dynamic_shape True --model_type rec
Export Success: output_folder/export_folder/abinet.mindir
Exporting model:master_resnet31, 2/24
output_folder/export_folder/master_resnet31.mindir exists and it will be overwritten if exported successfully.
Export Command: python export.py --model_name_or_config master_resnet31 --save_dir output_folder/export_folder --is_dynamic_shape True --model_type rec
Export Success: output_folder/export_folder/master_resnet31.mindir
Exporting model:cls_mobilenet_v3_small_100_model, 3/24
output_folder/export_folder/cls_mobilenet_v3_small_100_model.mindir exists and it will be overwritten if exported successfully.
Export Command: python export.py --model_name_or_config cls_mobilenet_v3_small_100_model --save_dir output_folder/export_folder --is_dynamic_shape True --model_type cls
Export Success: output_folder/export_folder/cls_mobilenet_v3_small_100_model.mindir
...
Exporting model:visionlan_resnet45, 23/24
output_folder/export_folder/visionlan_resnet45.mindir exists and it will be overwritten if exported successfully.
Export Command: python export.py --model_name_or_config visionlan_resnet45 --save_dir output_folder/export_folder --is_dynamic_shape True --model_type rec
Export Success: output_folder/export_folder/visionlan_resnet45.mindir
Exporting model:robustscanner_resnet31, 24/24
output_folder/export_folder/robustscanner_resnet31.mindir exists and it will be overwritten if exported successfully.
Export Command: python export.py --model_name_or_config robustscanner_resnet31 --save_dir output_folder/export_folder --is_dynamic_shape True --model_type rec
Export Success: output_folder/export_folder/robustscanner_resnet31.mindir
+----------------------------------+----------------+
|       Export Dynamic Model       | Dynamic Export |
+----------------------------------+----------------+
|              abinet              |    Success     |
|         master_resnet31          |    Success     |
| cls_mobilenet_v3_small_100_model |    Success     |
|          crnn_resnet34           |    Success     |
|         crnn_resnet34_ch         |    Success     |
|            crnn_vgg7             |    Success     |
|        dbnet_mobilenetv3         |    Success     |
|          dbnet_resnet18          |    Success     |
|          dbnet_resnet50          |    Success     |
|       dbnet_resnet50_ch_en       |    Success     |
|         dbnetpp_resnet50         |    Success     |
|      dbnetpp_resnet50_ch_en      |    Success     |
|         east_mobilenetv3         |    Success     |
|          east_resnet50           |    Success     |
|         fcenet_resnet50          |     Failed     |
|        psenet_mobilenetv3        |    Success     |
|         psenet_resnet50          |    Success     |
|         psenet_resnet152         |    Success     |
|          rare_resnet34           |    Success     |
|         rare_resnet34_ch         |    Success     |
|            svtr_tiny             |    Success     |
|           svtr_tiny_ch           |    Success     |
|        visionlan_resnet45        |    Success     |
|      robustscanner_resnet31      |    Success     |
+----------------------------------+----------------+
Converting model:abinet, 1/24
Convert Command: converter_lite --fmk=MINDIR --modelFile=all_output/export_folder/abinet.mindir --outputFile=all_output/convert_folder_lite_dynamic --optimize=ascend_oriented --configFile=all_output/convert_folder/temp_config.txt
Convert Failed 

Converting model:master_resnet31, 2/24
Convert Command: converter_lite --fmk=MINDIR --modelFile=all_output/export_folder/master_resnet31.mindir --outputFile=all_output/conveer/master_resnet31_lite_dynamic --optimize=ascend_oriented --configFile=all_output/convert_folder/temp_config.txt
Convert Failed 

Converting model:cls_mobilenet_v3_small_100_model, 3/24
Convert Command: converter_lite --fmk=MINDIR --modelFile=all_output/export_folder/cls_mobilenet_v3_small_100_model.mindir --outputFiletput/convert_folder/cls_mobilenet_v3_small_100_model_lite_dynamic --optimize=ascend_oriented --configFile=all_output/convert_folder/teig.txt
Convert Success: all_output/convert_folder/cls_mobilenet_v3_small_100_model_lite_dynamic.mindir
Benchmark Command: benchmark --modelFile=all_output/convert_folder/cls_mobilenet_v3_small_100_model_lite_dynamic.mindir --device=AscenutShapes=1,3,48,192 --loopCount=100 --warmUpLoopCount=10
Infer Dynamic Shape Success 
...
Converting model:visionlan_resnet45, 23/24
Convert Command: converter_lite --fmk=MINDIR --modelFile=all_output/export_folder/visionlan_resnet45.mindir --outputFile=all_output/coolder/visionlan_resnet45_lite_dynamic --optimize=ascend_oriented --configFile=all_output/convert_folder/temp_config.txt
Convert Failed 

Converting model:robustscanner_resnet31, 24/24
Convert Command: converter_lite --fmk=MINDIR --modelFile=all_output/export_folder/robustscanner_resnet31.mindir --outputFile=all_outpurt_folder/robustscanner_resnet31_lite_dynamic --optimize=ascend_oriented --configFile=all_output/convert_folder/temp_config.txt
Convert Failed 

+--------------------------------------------+-----------------+---------------+---------------+
| Convert Dynamic Model From Exported Mindir | Dynamic Convert | Infer Shape 0 | Infer Shape 1 |
+--------------------------------------------+-----------------+---------------+---------------+
|                   abinet                   |      Failed     |               |               |
|              master_resnet31               |      Failed     |               |               |
|      cls_mobilenet_v3_small_100_model      |     Success     |   1,3,48,192  |               |
|                 crnn_vgg7                  |     Success     |   1,3,32,100  |   1,3,32,132  |
|               crnn_resnet34                |     Success     |   1,3,32,100  |   1,3,32,132  |
|              crnn_resnet34_ch              |     Success     |   1,3,32,320  |   1,3,32,352  |
|             dbnet_mobilenetv3              |     Success     |  1,3,736,1280 |  1,3,704,1211 |
|               dbnet_resnet18               |     Success     |  1,3,736,1280 |  1,3,704,1211 |
|               dbnet_resnet50               |     Success     |  1,3,736,1280 |  1,3,704,1211 |
|            dbnet_resnet50_ch_en            |     Success     |  1,3,736,1280 |  1,3,704,1211 |
|              dbnetpp_resnet50              |     Success     | 1,3,1152,2048 | 1,3,1120,2080 |
|           dbnetpp_resnet50_ch_en           |     Success     | 1,3,1152,2048 | 1,3,1120,2080 |
|              east_mobilenetv3              |     Success     |  1,3,720,1280 |  1,3,816,1211 |
|               east_resnet50                |     Success     |  1,3,720,1280 |  1,3,816,1211 |
|              fcenet_resnet50               |      Failed     |               |               |
|             psenet_mobilenetv3             |     Success     |  1,3,736,1280 |  1,3,704,1211 |
|              psenet_resnet50               |     Success     |  1,3,736,1280 |  1,3,704,1211 |
|              psenet_resnet152              |     Success     | 1,3,1472,2624 | 1,3,1376,1307 |
|               rare_resnet34                |     Success     |   1,3,32,100  |   1,3,32,132  |
|              rare_resnet34_ch              |     Success     |   1,3,32,320  |   1,3,32,352  |
|                 svtr_tiny                  |     Success     |   1,3,64,256  |   1,3,64,288  |
|                svtr_tiny_ch                |     Success     |   1,3,32,320  |   1,3,32,352  |
|             visionlan_resnet45             |      Failed     |               |               |
|           robustscanner_resnet31           |      Failed     |               |               |
+--------------------------------------------+-----------------+---------------+---------------+

5. xxx

## Test Plan

(How should this PR be tested? Do you require special setup to run the test or repro the fixed bug?)

## Related Issues and PRs

(Is this PR part of a group of changes? Link the other relevant PRs and Issues here. Use https://help.github.com/en/articles/closing-issues-using-keywords for help on GitHub syntax)

…tools_bash

panshaowu · 2024-01-04T01:57:14Z

tools/data_for_export_convert.py

+        "infer_shape_list": ["1,3,736,1280", "1,3,704,1211"],
+        "var": "args0",
+    },
+    # "dbnet_resnet50_ch_en": {


此处注释是存在bug不支持吗？是否需要加TODO信息说明一下缘由？

panshaowu · 2024-01-04T01:57:24Z

tools/data_for_export_convert.py

+        "infer_shape_list": ["1,3,1152,2048", "1,3,1120,2080"],
+        "var": "args0",
+    },
+    # "dbnetpp_resnet50_ch_en": {


此处注释是存在bug不支持吗？是否需要加TODO信息说明一下缘由？

此处由于缺少dbnetpp_resnet50_ch_en和dbnet_resnet50_ch_en的ckpt自动下载。疑似dbnet_resnet50和dbnetpp_resnet50错写为dbnet_resnet50_ch_en和dbnetpp_resnet50_ch_en的ckpt下载路径。请查看该PR中mindocr/models/det_dbnet.py的修改。修改之后该注释取消。

panshaowu · 2024-01-04T02:47:37Z

tests/test_export_convert.py

+
+if __name__ == "__main__":
+    # not support export on Ascend 310P
+    # test_export_static_model(output_path)


此处为何大部分都要注释掉：

如依赖硬件，可考虑用npu-smi info判断本地ascend设备型号后，执行对应用例；

或者用注释说明，用户应选用哪些用例。

新建完文件夹了以后最好删掉，可以参照下面的代码写个装饰器：

def vfl_data_test(func): @wraps(func) def wrap_test(*args, **kwargs): try: stop_fl_process() clean_fl_temp_files() mkdir("temp") mkdir("temp/leader") mkdir("temp/follower") generate_random_data() func(*args, **kwargs) except Exception: logger.error("VFL data test catch exception") raise finally: logger.info("VFl data test begin to clear") stop_fl_process() clean_fl_temp_files() logger.info("VFl data test end clear") return wrap_test

不知道咋调github缩进，参照下面那个链接吧
https://gitee.com/mindspore/federated/blob/master/tests/ut/python/tests/common.py

CI的启停机制是怎么样的，代码涉及网络和多进程，最好加上超时报错和退出时清空进程

CI的启停机制是怎么样的，代码涉及网络和多进程，最好加上超时报错和退出时清空进程

CI中执行tests/ut和tests/st中的用例，并不会执行该文件

非CI可以不用加超时报错或者退出清空，忽略我上面的那些评论

不涉及CI

panshaowu · 2024-01-04T02:56:18Z

tests/test_export_convert.py

+def test_convert_static_model_from_download_mindir(output_path):
+    convert_tool = "converter_lite"
+    benchmark_tool = "benchmark"
+    save_path = f"{output_path}/convert_static_from_download"


构造路径时，需要考虑OS差异，建议使用：

file_path = os.getcwd() + os.sep + file_name file_path = os.path.join(os.getcwd(), file_name)

panshaowu · 2024-01-04T03:07:32Z

tools/export_convert_tool.py

+
+
+if __name__ == "__main__":
+    output_path = "./output"


是否需要判断路径是否存在，然后尝试新建文件夹？

已更新为parser传入，并加入路径判断，与是否文件覆盖选项

panshaowu · 2024-01-04T03:07:54Z

tests/test_export_convert.py

+from tools.export_convert_tool import ExportDynamicModel  # noqa
+from tools.export_convert_tool import ExportStaticModel  # noqa
+
+output_path = "./output"


是否需要判断路径是否存在，然后尝试新建文件夹？

已更新为parser传入，并加入路径判断，与是否文件覆盖选项

chengtianwu · 2024-01-04T03:21:15Z

tests/test_export_convert.py

+
+if __name__ == "__main__":
+    # not support export on Ascend 310P
+    # test_export_static_model(output_path)


新建完文件夹了以后最好删掉，可以参照下面的代码写个装饰器：

def vfl_data_test(func): @wraps(func) def wrap_test(*args, **kwargs): try: stop_fl_process() clean_fl_temp_files() mkdir("temp") mkdir("temp/leader") mkdir("temp/follower") generate_random_data() func(*args, **kwargs) except Exception: logger.error("VFL data test catch exception") raise finally: logger.info("VFl data test begin to clear") stop_fl_process() clean_fl_temp_files() logger.info("VFl data test end clear") return wrap_test

chengtianwu · 2024-01-04T03:21:47Z

tests/test_export_convert.py

+
+if __name__ == "__main__":
+    # not support export on Ascend 310P
+    # test_export_static_model(output_path)


不知道咋调github缩进，参照下面那个链接吧
https://gitee.com/mindspore/federated/blob/master/tests/ut/python/tests/common.py

chengtianwu · 2024-01-04T03:25:21Z

tests/test_export_convert.py

+
+if __name__ == "__main__":
+    # not support export on Ascend 310P
+    # test_export_static_model(output_path)


CI的启停机制是怎么样的，代码涉及网络和多进程，最好加上超时报错和退出时清空进程

chengtianwu · 2024-01-04T03:26:46Z

tools/export_convert_tool.py

+    convert_tool = "converter_lite"  # path to converter_lite
+    benchmark_tool = "benchmark"  # path to benchmark
+    save_path = f"{output_path}/convert_static_from_exported"  # path to save converted static model
+    exported_path = "path/to/exported/path"  # path to exported model


exported_path是不是没用到

不涉及CI

chengtianwu · 2024-01-04T03:26:50Z

tools/export_convert_tool.py

+    convert_tool = "converter_lite"  # path to converter_lite
+    benchmark_tool = "benchmark"  # path to benchmark
+    save_path = f"{output_path}/convert_dynamic_from_exported"  # path to save converted dynamic model
+    exported_path = "path/to/exported/path"  # path to exported model


exported_path是不是没用到

chengtianwu · 2024-01-04T03:27:45Z

tools/export_convert_tool.py

+            self.log_handle.write(f"Export Success: {exported_model_path}")
+            return True
+        else:
+            print("\033[31mExport Failed \033[0m")


日志中添加失败原因

chengtianwu · 2024-01-04T03:30:59Z

tools/export_convert_tool.py

+
+        subprocess.call(f"echo Export Command: {command}".split(), stdout=self.log_handle, stderr=self.log_handle)
+        print(f"\033[34mExport Command\033[0m: {command}")
+        subprocess.call(command.split(), stdout=self.log_handle, stderr=self.log_handle)


这种方式启动好像是非阻塞的？最好加个锁，然后校验下导出文件

非CI可以不用加，忽略我上面的评论

chengtianwu · 2024-01-08T03:00:08Z

tests/test_export_convert.py

+
+if __name__ == "__main__":
+    # not support export on Ascend 310P
+    # test_export_static_model(output_path)


非CI可以不用加超时报错或者退出清空，忽略我上面的那些评论

chengtianwu · 2024-01-08T03:00:34Z

tools/export_convert_tool.py

+
+        subprocess.call(f"echo Export Command: {command}".split(), stdout=self.log_handle, stderr=self.log_handle)
+        print(f"\033[34mExport Command\033[0m: {command}")
+        subprocess.call(command.split(), stdout=self.log_handle, stderr=self.log_handle)


非CI可以不用加，忽略我上面的评论

horcham and others added 5 commits November 3, 2023 14:28

Add tools bash for model exporting and converting

099f3c5

Merge branch 'main' of https://github.com/mindspore-lab/mindocr into …

9ba1b2a

…tools_bash

convert test

0fd3fcf

convert test

3c9e2e3

convert test

b086dbf

horcham force-pushed the tools_bash branch from 37918c1 to b086dbf Compare December 21, 2023 12:32

produce python tool for export and convert

b68dc13

horcham changed the title ~~[WIP] 模型转换与导出工具与测试~~ [WIP] tools for model exporting and converting Dec 24, 2023

horcham force-pushed the tools_bash branch 8 times, most recently from ca54557 to 1c82279 Compare December 26, 2023 10:21

horcham changed the title ~~[WIP] tools for model exporting and converting~~ tools for model exporting and converting Dec 26, 2023

chengtianwu approved these changes Jan 4, 2024

View reviewed changes

panshaowu reviewed Jan 4, 2024

View reviewed changes

chengtianwu reviewed Jan 4, 2024

View reviewed changes

panshaowu approved these changes Jan 5, 2024

View reviewed changes

chengtianwu approved these changes Jan 8, 2024

View reviewed changes

horcham force-pushed the tools_bash branch 5 times, most recently from 633c97d to 6e2f016 Compare January 12, 2024 04:50

Update export convert tools

6e68c2d

horcham force-pushed the tools_bash branch 2 times, most recently from 482ce08 to 20e09ad Compare January 12, 2024 08:08

horcham force-pushed the tools_bash branch 7 times, most recently from 9d0996e to 770473d Compare January 15, 2024 06:12

export convert tools

9316578

horcham force-pushed the tools_bash branch from 770473d to 9316578 Compare January 15, 2024 11:22

panshaowu merged commit 4bca517 into mindspore-lab:main Jan 16, 2024

tools for model exporting and converting #639

tools for model exporting and converting #639

Uh oh!

Conversation

horcham commented Dec 21, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

horcham Jan 12, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

horcham commented Dec 21, 2023 •

edited

Loading

horcham Jan 12, 2024 •

edited

Loading