Skip to content

Model Module

degirum.model.Model

Bases: ABC

Model class. Handles whole inference lifecycle for a single model: input data preprocessing, inference, and postprocessing.

Note

You never construct model objects yourself -- instead you call degirum.zoo_manager.ZooManager.load_model method to create degirum.model.Model instances for you.

degirum.model.Model.custom_postprocessor: Optional[type] property writable

Custom postprocessor class. When not None, the object of this class is returned as inference result. Such custom postprocessor classes must be inherited from degirum.postprocessor.InferenceResults class.

degirum.model.Model.devices_available: list property abstractmethod

The list of inference device indices which can be used for model inference.

degirum.model.Model.devices_selected: list property writable

The list of inference device indices selected for model inference.

degirum.model.Model.eager_batch_size: int property writable

The size of the batch (number of consecutive frames before this model is switched to another model during batch predict) to be used by device scheduler when inferencing this model.

degirum.model.Model.frame_queue_depth: int property writable

The depth of the model prediction queue. When the queue size reaches this value, the next prediction call will block until there will be space in the queue.

degirum.model.Model.image_backend: str property writable

Graphical library (backend) to use for graphical tasks -- one of 'pil', 'opencv', 'auto'

'auto' means try OpenCV first, if not installed, try PIL.

degirum.model.Model.inference_timeout_s: float property writable

The maximum time in seconds to wait for inference result from the model.

degirum.model.Model.input_crop_percentage: float property writable

Percentage of image to crop around. Valid range: [0..1].

degirum.model.Model.input_image_format: str property writable

Defines the image format for model inputs of image type -- one of 'JPEG' or 'RAW'.

degirum.model.Model.input_letterbox_fill_color: tuple property writable

Image fill color in case of 'letterbox' padding (see degirum.model.Model.input_pad_method property for details).

3-element RGB tuple.

degirum.model.Model.input_numpy_colorspace: str property writable

Input image colorspace -- one of 'RGB', 'BGR', or 'auto'.

This parameter is used only to identify colorspace for NumPy arrays.

'auto' translates to 'BGR' for opencv backend, and to 'RGB' for pil backend.

degirum.model.Model.input_pad_method: str property writable

Input image pad method -- one of 'stretch', 'letterbox', 'crop-first', or 'crop-last'.

  • In case of 'stretch', the input image will be resized to the model input size without preserving aspect ratio.
  • In case of 'letterbox', the input image will be resized to the model input size preserving aspect ratio.
  • In case of 'crop-first', the input image will be cropped to input_crop_percentage around the center and then resized.
  • In the case of 'crop-last', if the model's input dimensions are square, the image is resized with its smaller side matching the model dimension, preserving the aspect ratio. If the dimensions are rectangular, the image is resized and stretched to fit the model's input dimensions. After resizing, the image is cropped to the model's input dimensions and aspect ratio based on the 'input_crop_percentage' property.

The voids will be filled with solid color specified by input_letterbox_fill_color property. In all cases degirum.model.Model.input_resize_method property specifies the algorithm for resizing.

degirum.model.Model.input_resize_method: str property writable

Input image resize method -- one of 'nearest', 'bilinear', 'area', 'bicubic', or 'lanczos'.

degirum.model.Model.label_dictionary: Dict[int, str] property abstractmethod

Get model class label dictionary.

Each dictionary element is key-value pair, where the key is the class ID and the value is the class label string.

degirum.model.Model.measure_time: bool property writable

Flag to enable measuring and collecting inference time statistics.

Call degirum.model.Model.time_stats to query accumulated inference time statistics.

degirum.model.Model.model_info property

Return model information object to provide read-only access to model parameters.

New deep copy is created each time.

degirum.model.Model.non_blocking_batch_predict property writable

Flag to control the behavior of the generator object returned by predict_batch() method.

  • When the flag is set to True, the generator accepts None from the inference input data iterator object (passed as data parameter): If None is returned, the model predict step is skipped for this iteration. Also, when no inference results are available in the result queue at this iteration, the generator yields None result.

  • When the flag is set to False (default value), the generator does not allow None to be returned from the inference input data iterator object: If None is returned, an exception is raised. Also, when no inference results are available in the result queue at this iteration, the generator continues to the next iteration of the input data iterator.

  • Setting this flag to True allows using predict_batch() generator in a non-blocking manner, assuming the design of input data iterator object is also non-blocking, i.e., returning None when no data is available instead of waiting for the data. Every next element request from the generator will not block the execution waiting for either input data or inference results, returning None when no results are available.

degirum.model.Model.output_confidence_threshold: float property writable

Confidence threshold used in inference result post-processing.

Valid range: [0..1].

Only objects with scores higher than this threshold are reported.

Note

For classification models if degirum.model.Model.output_top_k parameter is set to non-zero value, then it supersedes this threshold -- degirum.model.Model.output_top_k highest score classes are always reported.

degirum.model.Model.output_max_classes_per_detection: int property writable

Max Detections Per Class number used in inference result post-processing, and specifies the maximum number of highest probability classes per anchor to be processed during the non-max suppression process for fast algorithm.

Applicable only for detection models.

degirum.model.Model.output_max_detections: int property writable

Max Detection number used in inference result post-processing, and specifies the total maximum objects of number to be detected.

Applicable only for detection models.

degirum.model.Model.output_max_detections_per_class: int property writable

Max Detections Per Class number used in inference result post-processing, and specifies the maximum number of objects to keep during per class non-max suppression process for regular algorithm.

Applicable only for detection models.

degirum.model.Model.output_nms_threshold: float property writable

Non-Max Suppression (NMS) threshold used in inference result post-processing.

Valid range: [0..1].

Applicable only for models which utilize NMS algorithm.

degirum.model.Model.output_pose_threshold: float property writable

Pose detection threshold used in inference result post-processing.

Valid range: [0..1].

Applicable only for pose detection models.

degirum.model.Model.output_postprocess_type: str property writable

Inference result post-processing type.

You may set it to 'None' to bypass post-processing.

degirum.model.Model.output_top_k: float property writable

The number of classes with highest scores to report for classification models.

When set to 0, then report all classes with scores greater than degirum.model.Model.output_confidence_threshold.

degirum.model.Model.output_use_regular_nms: bool property writable

Use Regular NMS value used in inference result post-processing and specifies the algorithm to use for detection postprocessing.

If value is True, regular Non-Max suppression algorithm is used -- NMS is calculated for each class separately and after that all results are merged.

If value is False, fast Non-Max suppression algorithm is used -- NMS is calculated for all classes simultaneously.

degirum.model.Model.overlay_alpha: float property writable

Alpha-blend weight for inference results drawing on overlay image.

float number in range [0..1].

See degirum.postprocessor.InferenceResults.image_overlay for more details.

degirum.model.Model.overlay_color property writable

Color for inference results drawing on overlay image.

3-element RGB tuple or list of 3-element RGB tuples.

The overlay_color property is used to define the color to draw overlay details. In the case of a single RGB tuple, the corresponding color is used to draw all the overlay data: points, boxes, labels, segments, etc. In the case of a list of RGB tuples the behavior depends on the model type:

  • For classification models different colors from the list are used to draw labels of different classes.
  • For detection models different colors are used to draw labels and boxes of different classes.
  • For pose detection models different colors are used to draw keypoints of different persons.
  • For segmentation models different colors are used to highlight segments of different classes.

If the list size is less than the number of classes of the model, then overlay_color values are used cyclically, for example, for three-element list it will be overlay_color[0], then overlay_color[1], overlay_color[2], and again overlay_color[0].

The default value of overlay_color is a single RBG tuple of yellow color for all model types except segmentation models. For segmentation models it is the list of RGB tuples with the list size equal to the number of model classes. You can use degirum.model.Model.label_dictionary property to obtain a list of model classes. Each color is automatically assigned to look pretty and different from other colors in the list.

degirum.model.Model.overlay_font_scale: float property writable

Font scale for inference results drawing on overlay image.

float positive number.

See degirum.postprocessor.InferenceResults.image_overlay for more details.

degirum.model.Model.overlay_line_width: int property writable

Line width for inference results drawing on overlay image.

See degirum.postprocessor.InferenceResults.image_overlay for more details.

degirum.model.Model.overlay_show_labels: bool property writable

Flag to enable/disable drawing class labels on overlay image.

See degirum.postprocessor.InferenceResults.image_overlay for more details.

degirum.model.Model.overlay_show_probabilities: bool property writable

Flag to enable/disable drawing class probabilities on overlay image.

See degirum.postprocessor.InferenceResults.image_overlay for more details.

degirum.model.Model.save_model_image: bool property writable

Flag to enable/disable saving of model input image in inference results.

Model input image is the image converted to AI model input specifications as raw binary array.

degirum.model.Model.__call__(data)

Perform whole inference lifecycle: input data preprocessing, inference and postprocessing.

Same as degirum.model.Model.predict.

degirum.model.Model.__enter__()

Context manager enter handler.

degirum.model.Model.__exit__(exc_type, exc_val, exc_tb)

Context manager exit handler.

degirum.model.Model.__init__(model_name, model_params)

Constructor.

Note

You never construct model objects yourself -- instead you call degirum.zoo_manager.ZooManager.load_model method to create degirum.model.Model instances for you.

degirum.model.Model.predict(data)

Perform whole inference lifecycle: input data preprocessing, inference, and postprocessing.

Parameters:

Name Type Description Default
data any

Inference input data. Input data type depends on the model.

  • If the model expects image data, then the input data is either:

    • Input image path string.
    • NumPy 3D array of pixels in a form HWC. where color dimension is native to selected graphical backend (RGB for 'pil' and BGR for 'opencv' backend)
    • PIL.Image object (only for 'pil' backend).
  • If the model expects audio data, then the input data is NumPy 1D array with audio data samples.

  • If the model expects raw tensor data, then the input data is NumPy multidimensional array with shape matching model input.

  • In case of multi-input model a list of elements of the supported data type is expected.

required

Returns:

Type Description
InferenceResults

Inference result object, which allows you to access inference results as a dictionary or as an overlay image if it is supported by the model. For your convenience, all image coordinates in case of detection models are converted from model coordinates to original image coordinates.

degirum.model.Model.predict_batch(data)

Perform whole inference lifecycle for all objects in given iterator object (for example, list).

Such iterator object should return the same object types which regular degirum.model.Model.predict method accepts.

Parameters:

Name Type Description Default
data iterator

Inference input data iterator object such as list or generator function.

Each element returned by this iterator can be one of the following:

  • A single input data object, in case of single-input model.
  • A list of input data objects, in case of multi-input model.
  • A tuple containing a pair of input data object or a list of input data objects as a first element and frame info object as a second element of the tuple.

The input data object type depends on the model.

  • If the model expects image data, then the input data object is either:

    • Input image path string.
    • NumPy 3D array of pixels in a form HWC, where color dimension is native to selected graphical backend (RGB for 'pil' and BGR for 'opencv' backend).
    • PIL.Image object (only for 'pil' backend).
  • If the model expects audio data, then the input data object is NumPy 1D array with audio data samples.

  • If the model expects raw tensor data, then the input data object is NumPy multidimensional array with shape matching model input.

The frame info object is passed to the inference result object unchanged and can be accessed via info property of the inference result object.

required

Returns:

Type Description
Iterator[InferenceResults]

Generator object which iterates over inference result objects. This allows you directly using the result of degirum.model.Model.predict_batch in for loops.

Example
    for result in model.predict_batch(['image1.jpg','image2.jpg']):
        print(result)

degirum.model.Model.predict_dir(path, *, recursive=False, extensions=['.jpg', '.jpeg', '.png', '.bmp'])

Perform whole inference lifecycle for all files from specified directory matching given file extensions.

Supports only single-input models.

Parameters:

Name Type Description Default
path str

Directory name containing files to be processed.

required
recursive bool

True to recursively walk through all subdirectories in a directory. Default is False.

False
extensions list[str]

Single string or list of strings containing file extension(s) to process.

['.jpg', '.jpeg', '.png', '.bmp']

Returns:

Type Description
Iterator[InferenceResults]

Generator object to iterate over inference result objects. This allows you directly using the result of degirum.model.Model.predict_dir in for loops.

Example
    for result in model.predict_dir('./some_path'):
        print(result)

degirum.model.Model.reset_time_stats()

Reset inference time statistics.

degirum.model.Model.time_stats method will return empty dictionary after this call.

degirum.model.Model.time_stats()

Query inference time statistics.

Returns:

Type Description
dict

Dictionary containing time statistic objects.

  • A key in that dictionary is a string description of a particular inference step.
  • Each statistic object keeps min, max, and average values in milliseconds, accumulated over all inferences performed on this model since the model creation of last call of statistic reset method degirum.model.Model.reset_time_stats.
  • Time statistics are accumulated only when degirum.model.Model.measure_time property is set to True.