Base Classes for Postprocessing Plugins

A model postprocessing plugin is used to postprocess the output tensor of model inference and write the corresponding result to the metadata. The input and output types required for postprocessing of different inference job types are different. Therefore, different postprocessing plugins are required. To facilitate reuse, Vision SDK has extracted some APIs and members to the postprocessing base class. For each type of tasks, such as object detection, the postprocessing .so file is dynamically loaded to support polymorphism of various models (such as YOLOv3, Faster R-CNN, and SsdVgg-16).

Function	Performs postprocessing on the output tensor after model inference.
Constraints	The postprocessing plugins can be used after being connected to the mxpi_tensorinfer inference plugin. Only MxpiTensorPackageList can be used as the metadata input. The parent class does not generate plugins, and the subclass inherits the parent class to generate plugins.
Parent Class	MxModelPostProcessorBase and MxImagePostProcessorBase (image postprocessing base class, inherited from MxModelPostProcessorBase)
Input/Output	Input: buffer (data type: MxpiBuffer), metadata (data type: MxpiTensorPackageList) Output: buffer (data type: MxpiBuffer), metadata (data type: MxpiObjectList, MxpiClassList, MxpiImageMaskList, MxpiTextObjectList, etc.)
Port Format (Caps)	Static input: {"metadata/tensor"} Static output: subclass overriding
Property	For details, see Table 1.

**Table 1** Properties of MxModelPostProcessorBase and MxImagePostProcessorBase
Property Name	Description	Mandatory or Not	Modifiable or Not
deviceId	Ascend device ID, which is specified by the deviceId property in the stream_config field. You do not need to set the ID.	No	Yes
postProcessLibPath	Path of the .so file of the postprocessing DLL. If the path is not specified, the model inference result is directly written to the metadata MxpiTensorPackageList and output to the position specified by outputDeviceId.	Yes	Yes
labelPath	Path of the postprocessing class label	No	Yes
dataSource	Index of the input data (generally the name of the upstream element). The default value is the key value of the output port of the upstream plugin.	No	Yes
postProcessConfigPath	Path of the postprocessing configuration file	Yes	Yes
postProcessConfigContent	Postprocessing configuration	No	Yes
dataSourceResize	This property is available only in the subclass plugins that inherit MxImagePostProcessorBase (image postprocessing base class). Whether to map the coordinates in the postprocessing of the model back to the image before resizing. If this property is not set, the value is obtained from the upstream plugin of the inference plugin by default. If the value cannot be obtained, the coordinates are not scaled.	No	Yes
dataSourceRoiBoxes	This property is available only in the subclass plugins that inherit MxImagePostProcessorBase (image postprocessing base class). Specifies whether to map the coordinates of the model inference to the source image before cropping. If this property is not set, the image is not mapped to the source image by default. To map the image, set the property to the name of the image cropping plugin.	No	Yes
funcLanguage	Development language of the postprocessing plugin, that is, C++ or Python	No	Yes
className	Postprocessing class name	Yes	Yes
pythonModule	Name of the loaded postprocessing module, which is the same as the module name in Python or .so file	Yes	Yes
dataSourceImage	It is under internal debugging. Do not use it.	No	No

**Table 2** Python postprocessing properties
Property Name	Description	Mandatory or Not	Modifiable or Not
funcLanguage	Development language of the postprocessing plugin, that is, C++ or Python	No	Yes
postProcessLibPath	Directory of the postprocessing DLL .so file. This directory contains the postprocessing Python file or .so file	Yes	Yes
className	Postprocessing class name	Yes	Yes
pythonModule	Name of the loaded postprocessing module, which is the same as the module name in Python or .so file	Yes	Yes
labelPath	Path of the postprocessing class label	No	Yes
dataSource	Index of the input data (generally the name of the upstream element).	Yes	Yes
postProcessConfigPath	Path of the postprocessing configuration file	Yes	Yes
deviceId	Ascend device ID, which is specified by the deviceId property in the stream_config field. You do not need to set the ID.	No	Yes
dataSourceResize	This property is available only in the subclass plugins that inherit MxImagePostProcessorBase (image postprocessing base class). Whether to map the coordinates in the postprocessing of the model back to the image before resizing. If this property is not set, the value is obtained from the upstream plugin of the inference plugin by default. If the value cannot be obtained, the coordinates are not scaled.	No	Yes
dataSourceRoiBoxes	This property is available only in the subclass plugins that inherit MxImagePostProcessorBase (image postprocessing base class). Specifies whether to map the coordinates of the model inference to the source image before cropping. If this property is not set, the image is not mapped to the source image by default. To map the image, set the property to the name of the image cropping plugin.	No	Yes

Figure 1 Inheritance relationship of postprocessing plugin classes

Parent topic: Model Postprocessing Plugins