Model configuration#

Model’s static method create_model() has two overloads. One constructs the model from a string (a path or a model name) and the other takes an already constructed InferenceAdapter. The first overload configures a created model with values taken from configuration dict function argument and from model’s intermediate representation (IR) stored in .xml in model_info section of rt_info. Values provided in configuration have priority over values in IR rt_info. If no value is specified in configuration nor in rt_info the default value for a model wrapper is used. For Python configuration values are accessible as model wrapper member fields.

List of values#

The list features only model wrappers which introduce new configuration values in their hierarchy.

  1. model_type: str - name of a model wrapper to be created

  2. layout: str - layout of input data in the format: “input0:NCHW,input1:NC”

ImageModel and its subclasses#

  1. mean_values: List - normalization values, which will be subtracted from image channels for image-input layer during preprocessing

  2. scale_values: List - normalization values, which will divide the image channels for image-input layer

  3. reverse_input_channels: bool - reverse the input channel order

  4. resize_type: str - crop, standard, fit_to_window or fit_to_window_letterbox

  5. embedded_processing: bool - flag that pre/postprocessing embedded

  6. pad_value: int - pad value for resize_image_letterbox embedded into a model

  7. orig_height: int - model original image height before embedded preprocessing

  8. orig_width: int - model original image width before embedded preprocessing

AnomalyDetection#

  1. image_threshold: float - Image threshold that is used for classifying an image as anomalous

  2. pixel_threshold: float - Pixel level threshold used to segment anomalous regions in the image

  3. normalization_scale: float - Scale by which the outputs are divided. Used to apply min-max normalization

  4. task: str - Outputs segmentation masks, bounding boxes, or anomaly score based on the task type

  5. labels: List - list of class labels

  6. path_to_labels: str - path to file with labels. Overrides the labels, if they sets via ‘labels’ parameter

ClassificationModel#

  1. topk: int - number of most likely labels

  2. labels: List - list of class labels

  3. path_to_labels: str - path to file with labels. Overrides the labels, if they sets via ‘labels’ parameter

  4. multilabel: bool - predict a set of labels per image

  5. hierarchical: bool - predict a hierarchy of labels per image hierarchical_config

  6. hierarchical_config: str - a serialized configuration for decoding hierarchical predictions

  7. confidence_threshold: float - probability threshold value for multilabel or hierarchical predictions filtering

  8. hierarchical_postproc: str - greedy (default) or probabilistic postprocessing for hierarchical predictions

  9. confidence_threshold: float - probability threshold value for bounding box filtering

  10. output_raw_scores: bool - output all scores for multiclass classification

DetectionModel and its subclasses#

  1. confidence_threshold: float - probability threshold value for bounding box filtering

  2. labels: List - List of class labels

  3. path_to_labels: str - path to file with labels. Overrides the labels, if they sets via labels parameter

YOLO and its subclasses#

  1. iou_threshold: float - threshold for non-maximum suppression (NMS) intersection over union (IOU) filtering

YoloV4#

  1. anchors: List - list of custom anchor values

  2. masks: List - list of mask, applied to anchors for each output layer

YOLOv5, YOLOv8, YOLO11#

  1. agnostic_nms: bool - if True, the model is agnostic to the number of classes, and all classes are considered as one

  2. iou_threshold: float - threshold for non-maximum suppression (NMS) intersection over union (IOU) filtering

YOLOX#

  1. iou_threshold: float - threshold for non-maximum suppression (NMS) intersection over union (IOU) filtering

MaskRCNNModel#

  1. confidence_threshold: float - probability threshold value for bounding box filtering

  2. labels: List - list of class labels

  3. path_to_labels: str - path to file with labels. Overrides the labels, if they sets via labels parameter

  4. postprocess_semantic_masks: bool - resize and apply 0.5 threshold to instance segmentation masks

SegmentationModel and its subclasses#

  1. labels: List - list of class labels

  2. path_to_labels: str - path to file with labels. Overrides the labels, if they sets via ‘labels’ parameter

  3. blur_strength: int - blurring kernel size. -1 value means no blurring and no soft_threshold

  4. soft_threshold: float - probability threshold value for bounding box filtering. inf value means no blurring and no soft_threshold

  5. return_soft_prediction: bool - return raw resized model prediction in addition to processed one

ActionClassificationModel#

  1. labels: List - list of class labels

  2. path_to_labels: str - path to file with labels. Overrides the labels, if they sets via ‘labels’ parameter

  3. resize_type: str - crop, standard, fit_to_window or fit_to_window_letterbox

  4. pad_value: int - pad value for resize_image_letterbox embedded into a model

  5. orig_height: int - model original image height before embedded preprocessing

  6. orig_width: int - model original image width before embedded preprocessing

  7. embedded_processing: bool - flag that pre/postprocessing embedded

  8. mean_values: List - normalization values, which will be subtracted from image channels for image-input layer during preprocessing

  9. scale_values: List - normalization values, which will divide the image channels for image-input layer

  10. reverse_input_channels: bool - reverse the input channel order

NOTE ActionClassificationModel isn’t subclass of ImageModel.