調試open_model_zoo/mask_rcnn_demo
接前面一篇,編譯好了DEMO,我們繼續玩轉OpenVINO,試一下open_model_zoo中的模型。
假設你要調試open_model_zoo中的某個模型(注意,下面的命令中,使用你自己想用的模型,我一般是好幾個同時轉換,完了隨機測試的)。
先是要完成Optimizer工作,轉換模型得到IR文件,
命令如下
python mo_tf.py --input_model
E:/mask_rcnn_resnet50_atrous_coco_2018_01_28/frozen_inference_graph.pb
--tensorflow_use_custom_operations_config extensions/front/tf/mask_rcnn_support.json
--tensorflow_object_detection_api_pipeline_config E:/mask_rcnn_inception_resnet_v2_atrous_coco_2018_01_28/pipeline.config
喜歡用vscode調試的朋友可以看下面的launch.json文件,
{
// Use IntelliSense to learn about possible attributes.
// Hover to view descriptions of existing attributes.
// For more information, visit: https://go.microsoft.com/fwlink/?linkid=830387
"version": "0.2.0",
"configurations": [
{
"name": "Python: 當前文件",
"type": "python",
"request": "launch",
"program": "${file}",
"console": "integratedTerminal",
"justMyCode": false,
"args": [
"--input_model","E:\\mask_rcnn_resnet50_atrous_coco_2018_01_28\\frozen_inference_graph.pb",
"--tensorflow_use_custom_operations_config","D:/devOpenVino/openvino_2020.3.194/deployment_tools/model_optimizer/extensions/front/tf/mask_rcnn_support.json",
"--tensorflow_object_detection_api_pipeline_config","E:/mask_rcnn_inception_resnet_v2_atrous_coco_2018_01_28/pipeline.config"
]
}
]
}
如果你和我一樣,沒有指定輸出文件名,轉換完成後得到的都是frozen_inference_graph.bin和frozen_inference_graph.xml文件,要注意改成相應的模型文件名,否則多了就弄混了。
下面,我們開始用VS2019中的C++ Demo來測試這些模型。這些DEMO在我們上一講《玩轉OpenVINO_cpp samples的編譯》中已經編譯好了,現在拿來用。
添加路徑一
如果你使用debug版本,那麼環境變量path路徑設置中添加
C:\IntelSWTools\openvino_2020.3.194\deployment_tools\inference_engine\bin\intel64\Debug
同時,把opencv_world430d.dll拷貝到該文件夾下面(不想拷貝的話,自己添加路徑也行,反正就是讓程序能找到這個dll文件)
如果是release版本,則添加
C:\IntelSWTools\openvino_2020.3.194\deployment_tools\inference_engine\bin\intel64\Release,
同時,把opencv_world430.dll拷貝到該文件夾下面
總的來說,這裏有不少dll文件是intel ineference_engine要用到的。
添加路徑二
還有一些路徑也是必須添加的,
C:\IntelSWTools\openvino_2020.3.194\deployment_tools\inference_engine\external\tbb\bin
C:\IntelSWTools\openvino_2020.3.194\deployment_tools\ngraph\lib
調試運行
運行的項目名稱是mask_rcnn_demo。
具體可參考:https://docs.openvinotoolkit.org/latest/_demos_mask_rcnn_demo_README.html
我把說明摘錄一部分如下(注意:這裏是linux下的格式,我後面說明中用到的是windows系統中的格式,在命令使用上有點小小的差異)
./mask_rcnn_demo -h
InferenceEngine:
API version ............ <version>
Build .................. <number>
mask_rcnn_demo [OPTION]
Options:
-h Print a usage message.
-i "<path>" Required. Path to a .bmp image.
-m "<path>" Required. Path to an .xml file with a trained model.
-l "<absolute_path>" Required for CPU custom layers. Absolute path to a shared library with the kernels implementations.
Or
-c "<absolute_path>" Required for GPU custom kernels. Absolute path to the .xml file with the kernels descriptions.
-d "<device>" Optional. Specify the target device to infer on (the list of available devices is shown below). Use "-d HETERO:<comma-separated_devices_list>" format to specify HETERO plugin. The demo will look for a suitable plugin for a specified device (CPU by default)
-detection_output_name "<string>" Optional. The name of detection output layer. Default value is "reshape_do_2d"
-masks_name "<string>" Optional. The name of masks layer. Default value is "masks"
查看幫助文檔: mask_rcnn_demo --h
C:\IntelSWTools\openvino_2020.3.194\deployment_tools\open_model_zoo\demos\dev\intel64\Debug>mask_rcnn_demo --h
InferenceEngine: 00007FFCC7C49BC8
mask_rcnn_demo [OPTION]
Options:
-h Print a usage message.
-i "<path>" Required. Path to a .bmp image.
-m "<path>" Required. Path to an .xml file with a trained model.
-l "<absolute_path>" Required for CPU custom layers. Absolute path to a shared library with the kernels implementations.
Or
-c "<absolute_path>" Required for GPU custom kernels. Absolute path to the .xml file with the kernels descriptions.
-d "<device>" Optional. Specify the target device to infer on (the list of available devices is shown below). Use "-d HETERO:<comma-separated_devices_list>" format to specify HETERO plugin. The demo will look for a suitable plugin for a specified device (CPU by default)
-detection_output_name "<string>" Optional. The name of detection output layer. Default value is "reshape_do_2d"
-masks_name "<string>" Optional. The name of masks layer. Default value is "masks"
Available target devices: CPU GNA
這裏圖片必須是bmp格式。
如何輸入圖片地址呢?官方給出的命令如下,
./mask_rcnn_demo -i <path_to_image>/inputImage.bmp -m <path_to_model>/mask_rcnn_inception_resnet_v2_atrous_coco.xml
事實上,用命令行輸入的方式, OpenVINO中由一個叫args_helper.hpp的文件來處理,其中一段的代碼如下,
/**
* @brief This function find -i/--images key in input args
* It's necessary to process multiple values for single key
* @return files updated vector of verified input files
*/
inline void parseInputFilesArguments(std::vector<std::string> &files) {
std::vector<std::string> args = gflags::GetArgvs();
bool readArguments = false;
for (size_t i = 0; i < args.size(); i++) {
if (args.at(i) == "-i" || args.at(i) == "--images") {
readArguments = true;
continue;
}
if (!readArguments) {
continue;
}
if (args.at(i).c_str()[0] == '-') {
break;
}
readInputFilesArguments(files, args.at(i));
}
}
就是說,輸入圖片的格式爲以下兩者都可以,
-i xyz.bmp 或者 --images <沒仔細研究,這裏是要文件夾吧還是xyz.bmp>
在VS2019中調試運行的話,直接把項目的調試參數設置爲上述格式即可,例如,
-i J:\BigData\default.bmp -m E:\mask_rcnn_resnet50_atrous_coco_2018_01_28\frozen_inference_graph.xml
我用純CPU試了一下這個DEBUG模式,超級慢啊!在Release模式下,隨便找了一張圖,大約也花了好幾秒,感覺不出來哪裏加快了。
當然,要琢磨的地方還很多,這裏暫不涉及這些細節了,先玩起來吧。