Classification Unit

Savant supports classification models. The element type “Classification Unit” is used for inferring classifier models. Classifiers are applied to objects existing in the frame, whether they are objects previously detected by a detector or a default meta-object that represents the entire frame. The result of the classifier’s work is meta-information about the assigned class attached to the target objects as an attribute.

Configuration Parameters Specific to The Classification Model Unit

The classifier’s inference results in a list of attributes appended to the object’s metadata. This list is configured in the element.model.output.attributes section.

For each of the output layers of the classifier, you must specify a separate attribute. By default, the correspondence of output layers to attributes is determined by the enumeration order; however, when using a custom converter, this correspondence is determined by the logic of the converter.

The mandatory parameter when configuring an attribute is name: a value that defines the metadata name for the attribute, which can be used later to obtain its value.

The optional parameters when configuring an attribute are:

  • labels defines a dictionary that sets symbolic values for class indices obtained after inferring the model. It translates 0, 1, … into cat, dog, etc. If this dictionary is not defined, then attributes will be used with numeric values.

  • threshold is a float number that sets a confidence value threshold for the attribute to be added to the results. The attribute will be added to metadata only if the associated confidence value is strictly above the specified threshold. If it is not defined, the attribute is added to results unconditionally.

  • multi_label is a flag that specifies if every class that passes the configured threshold should be inlcuded in results. By default, it is set to false, so only the class with the highest confidence value will be inlcuded in results.

  • internal is a flag that makes the classes be ignored when metadata is passed to the pipeline output. By default, it is false. If true, the attribute will only be available within the pipeline. This parameter should be used if the attribute is auxiliary and meaningless outside the pipeline.

To use a custom model, you may need to write a result converter: pyfunc, which parses the output tensors and builds the list of attribute values. The converter is specified in the element.model.output.converter section:

  • module - a value that defines the Python module containing the converter code.

  • class_name - a value that specifies the class from the module. The converter class must be a derived from BaseAttributeModelOutputConverter.

It is important to note that when element.model.output.converter is specified, the classifier will be configured as a custom Deepstream model with the output in the form of a raw tensor. In this case all optional parameters described above will be used.

If the converter is not specified, the model will be configured as a regular nvinfer classifier. In this case, the optional configuration parameters for the attributes listed above (except threshold) are not used, and custom parsing of classifier results must be done with Deepstream functionality by defining the element.model.custom_lib_path and element.model.parse_classifier_func_name parameters in the configuration. If a custom function is not specified, Deepstream uses the default softmax layer function to parse the classification results.


The converter can access data with CPU-based tools (NumPy) or in GPU with CuPy. Typically, NumPy-based processing is a default choice; however, if you have a lot of data to process, you may consider using CuPy to speed up the processing.

In addition, for regular Deepstream classifiers, it is impossible to set a minimum confidence threshold for each attribute individually, so the first specified threshold value will be applied to every attribute.

- element: nvinfer@classifier
  name: Secondary_CarColor
    format: caffe
    model_file: resnet18.caffemodel
    mean_file: mean.ppm
    label_file: labels.txt
    batch_size: 16
      object: Primary_Detector.Car
      object_min_width: 64
      object_min_height: 64
      color_format: bgr
      layer_names: [predictions/Softmax]
        - name: car_color
          threshold: 0.51