Compound Models

DeGirum Tools API Reference Guide. Chain multiple models in cropping, combining or async flows.

This API Reference is based on DeGirum Tools version 0.18.0.

Compound Models Module Overview

This module provides a toolkit for creating compound models using the DeGirum PySDK.

A compound model orchestrates multiple underlying models into a pipeline to enable complex inference scenarios. Common examples include:

• Detecting objects and then classifying each detected object.

• Running coarse detection first, then applying a refined detection model on specific regions.

• Combining outputs from multiple independent models into a unified inference result.

Compound models run in a single thread and are intended primarily for simple usage scenarios. Compound models still provide efficient batch prediction pipelining using batch_predict() in non-blocking mode. For more performant applications requiring better scalability and more flexible connections, we recommend using Gizmos, which in multiple threads.

Key Concepts

• Model Composition: Compound models sequentially (or concurrently) invoke multiple models. Typically, results from the first model (e.g., bounding boxes from detection) feed into subsequent models (classification or refined detection).

• Pipeline Workflow: A typical workflow involves:

  1. Run model1 to identify regions of interest (ROIs).

  2. Crop these ROIs and run them through model2.

  3. Integrate or transform outputs from model2 back into the original context.

• Unified Model Interface: All compound models follow the same interface as regular models in DeGirum SDK, providing .predict() for single frames and .predict_batch() for iterators of frames.

Included Compound Models

• CombiningCompoundModel: Combines results from two models run concurrently on the same input.

• CroppingCompoundModel: Crops regions identified by model1 and feeds them into model2.

• CroppingAndClassifyingCompoundModel: Specialized pipeline: object detection (model1) followed by classification (model2) of each detected object.

• CroppingAndDetectingCompoundModel: Pipeline for refined detection: initial coarse detection (model1) followed by detailed detection (model2) within each ROI.

• RegionExtractionPseudoModel: Extracts predefined regions of interest without actual inference, optionally filtering by motion detection.

Basic Usage Examples

Detection + Classification:

import degirum as dg
from degirum_tools.compound_models import CroppingAndClassifyingCompoundModel

# Declaring variables
your_detection_model = "yolov8n_relu6_coco--640x640_quant_n2x_orca1_1"
your_classification_model = "resnet50_imagenet--224x224_pruned_quant_n2x_orca1_1"
your_host_address = dg.CLOUD  # Can be dg.CLOUD, host:port, or dg.LOCAL
your_model_zoo = "degirum/degirum"
your_token = "<token>"

# Specify images for inference
your_image = "path/image1.jpg"
your_images = ["path/image2.jpg", "path/image3.jpg"]

# Loading the individual models
detector = dg.load_model(
    model_name=your_detection_model,
    inference_host_address=your_host_address,
    zoo_url=your_model_zoo,
    token=your_token
)

classifier = dg.load_model(
    model_name=your_classification_model,
    inference_host_address=your_host_address,
    zoo_url=your_model_zoo,
    token=your_token
)

# Creating a compound model pipeline
compound_model = CroppingAndClassifyingCompoundModel(detector, classifier)

# Single frame inference using predict()
print("Using predict():")
single_result = compound_model(your_image)
print(single_result)

# Batch inference using predict_batch()
print("Using predict_batch():")
for batch_result in compound_model.predict_batch(your_images):
    print(batch_result)

Detection + Detection:

import degirum as dg
from degirum_tools.compound_models import CombiningCompoundModel

# Declaring variables
your_detection_model_1 = "yolov8n_relu6_car--640x640_quant_n2x_orca1_1"
your_detection_model_2 = "yolov8n_relu6_coco--640x640_quant_n2x_orca1_1"
your_host_address = dg.CLOUD  # Can be dg.CLOUD, host:port, or dg.LOCAL
your_model_zoo = "degirum/degirum"
your_token = "<token>"

# Specify images for inference
your_image = "path/image1.jpg"
your_images = ["path/image2.jpg", "path/image3.jpg"]

# Loading the individual detection models
detector1 = dg.load_model(
    model_name=your_detection_model_1,
    inference_host_address=your_host_address,
    zoo_url=your_model_zoo,
    token=your_token
)

detector2 = dg.load_model(
    model_name=your_detection_model_2,
    inference_host_address=your_host_address,
    zoo_url=your_model_zoo,
    token=your_token
)

# Creating a compound model that merges results from both detectors
compound_detector = CombiningCompoundModel(detector1, detector2)

# Single frame inference using predict()
print("Using predict():")
single_result = compound_detector(your_image)
print(single_result.results)

# Batch inference using predict_batch()
print("Using predict_batch():")
for batch_result in compound_detector.predict_batch(your_images):
    print(batch_result.results)

See class-level documentation below for details on individual classes and additional configuration options.

Classes

ModelLike

ModelLike

Bases: ABC

A base class which provides a common interface for all models, similar to PySDK model class.

When calling predict_batch(data), each item in data can be:

• A single frame (image/array/etc.), or

• A 2-element tuple in the form (frame, frame_info).

The frame_info object (of any type) then appears in the final InferenceResults.info` attribute, allowing you to carry custom metadata through the pipeline.

Functions

__call__(data)

__call__(data)

Perform a whole inference lifecycle on a single frame (callable alias to predict()).

Parameters:

Name	Type	Description	Default
data	any	Inference input data, typically an image or array, or a tuple (frame, frame_info).	required

Returns:

Type	Description
InferenceResults or None	The combined inference result object, or None if no result.

predict(data)

predict(data)

Perform a whole inference lifecycle on a single frame.

Parameters:

Name	Type	Description	Default
data	any	Inference input data, typically an image or array, or a tuple (frame, frame_info).	required

Returns:

Type	Description
InferenceResults or None	The combined inference result object, or None if no result.

predict_batch(data)

predict_batch(data)

abstractmethod

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

Each item in data can be a single frame (any type acceptable to the model) or a 2-element tuple (frame, frame_info). In the latter case, frame_info is carried through and placed in InferenceResults.info` for that frame.

Parameters:

Name	Type	Description	Default
data	iterator	Inference input data iterator object such as a list or a generator function. Each element returned by this iterator should be compatible with what regular PySDK models accept.	required

Returns:

Type	Description
Iterator[InferenceResults or None]	A generator or iterator over the inference result objects (or None in non-blocking mode). This allows you to use the result in for loops.

FrameInfo

FrameInfo

Class to hold frame info.

By default, DeGirum PySDK allows you to pass any arbitrary object as 'frame info' alongside each frame in predict_batch().

Attributes:

Name	Type	Description
result1	any	The result object produced by the first model in a compound pipeline. For instance, an InferenceResults object.
sub_result	int	The index of a sub-result within result1 (e.g., which bounding box led to this cropped image).

CompoundModelBase

CompoundModelBase

Bases: ModelLike

Compound model class which combines two models into one pipeline.

One model is considered primary (model1), and the other is nested (model2).

The primary model (model1) processes the input frames. Its results are then passed to the nested model (model2).

Attributes

non_blocking_batch_predict

non_blocking_batch_predict

property writable

Flag controlling whether predict_batch() operates in non-blocking mode for model1. In non-blocking mode, predict_batch() can yield None when no results are immediately available.

Returns:

Type	Description
bool	True if non-blocking mode is enabled, False otherwise.

Classes

NonBlockingQueue

NonBlockingQueue

Bases: Queue

Specialized non-blocking queue which acts as an iterator to feed data to the nested model.

Functions

__iter__

__iter__()

Yield items from the queue until a None sentinel is reached.

Yields:

Type	Description
any or None	The item from the queue, or None if the queue is empty.

Functions

__getattr__(attr)

__getattr__(attr)

Fallback for getters of model-like attributes to the primary model (model1).

__init__(model1, ...)

__init__(model1, model2)

Constructor.

Parameters:

Name	Type	Description	Default
model1	ModelLike	Model to be used for the first step of the pipeline.	required
model2	ModelLike	Model to be used for the second step of the pipeline.	required

__setattr__(key, ...)

__setattr__(key, value)

Intercepts attempts to set attributes. If the attribute already exists on the instance, the class, or is being set inside __init__, the attribute is set normally. Otherwise, the attribute assignment is delegated to the primary model (model1) if defined. This prevents adding new attributes outside of __init__.

attach_analyzers(analyzers)

attach_analyzers(analyzers)

Attach analyzers to a model.

Parameters:

Name	Type	Description	Default
analyzers	Union[ResultAnalyzerBase, list[ResultAnalyzerBase], None]	A single analyzer, or a list of analyzer objects, or None to detach all analyzers.	required

predict_batch(data)

predict_batch(data)

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

Works in a pipeline fashion

1. Pass input frames (or (frame, frame_info) tuples) to model1.

2. Use queue_result1(result1) to feed model2.

3. Collect model2 results, transform them with transform_result2(result2),

4. Yield the final output.

Parameters:

Name	Type	Description	Default
data	iterator	Inference input data iterator object such as a list or a generator function. Each element returned should be compatible with model inference requirements.	required

Returns:

Type	Description
Iterator[InferenceResults or None]	Generator object which iterates over the combined inference result objects (or None in non-blocking mode). This allows you to use the result in for loops.

queue_result1(result1)

queue_result1(result1)

abstractmethod

Process the result of the first model and put it into the queue.

Parameters:

Name	Type	Description	Default
result1	InferenceResults	Prediction result of the first model.	required

transform_result2(result2)

transform_result2(result2)

abstractmethod

Transform (or integrate) the result of the second model.

Parameters:

Name	Type	Description	Default
result2	InferenceResults	Prediction result of the second model.	required

Returns:

Type	Description
InferenceResults or None	Transformed/combined result to be returned by the compound model. If None, that means no result is produced at this iteration.

CombiningCompoundModel

CombiningCompoundModel

Bases: CompoundModelBase

Compound model class which executes two models in parallel on the same input data and merges their results.

Restriction: both models should produce the same type of inference results (e.g., both detection).

Functions

queue_result1(result1)

queue_result1(result1)

Queues the original image from result1 and a new FrameInfo instance that references result1. This (frame, frame_info) tuple is then read by model2.

Parameters:

Name	Type	Description	Default
result1	InferenceResults	Inference result from model1. We extract result1.image as the frame, and create a FrameInfo so we know which result1 this frame corresponds to.	required

transform_result2(result2)

transform_result2(result2)

Merges results from model2 into result1 that was stored in FrameInfo.

This implementation appends the second model's inference results to the first model's result list.

Parameters:

Name	Type	Description	Default
result2	InferenceResults	Inference result of the second model, which has info attribute containing the FrameInfo.	required

Returns:

Type	Description
InferenceResults	The merged inference results (model1 + model2).

CroppingCompoundModel

CroppingCompoundModel

Bases: CompoundModelBase

Compound model class which crops the original image according to results of the first model and then passes these cropped images to the second model.

Restriction: the first model should be of object detection type.

Functions

__init__(model1, ...)

__init__(model1, model2, crop_extent=0.0, crop_extent_option=CropExtentOptions.ASPECT_RATIO_NO_ADJUSTMENT)

Constructor.

Parameters:

Name	Type	Description	Default
model1	ModelLike	Object detection model that produces bounding boxes.	required
model2	ModelLike	Classification model that will process each cropped region.	required
crop_extent	float	Extent of cropping (in percent of bbox size) to expand the bbox.	0.0
crop_extent_option	CropExtentOptions	Method of applying extended crop to the input image for model2.	ASPECT_RATIO_NO_ADJUSTMENT

queue_result1(result1)

queue_result1(result1)

Put the original image into the queue, along with bounding boxes from the first model.

If no bounding boxes are detected, puts a small black image to keep the pipeline in sync.

Parameters:

Name	Type	Description	Default
result1	InferenceResults	Prediction result of the first (object detection) model.	required

CroppingAndClassifyingCompoundModel

CroppingAndClassifyingCompoundModel

Bases: CroppingCompoundModel

Compound model class which

1. Runs an object detection (model1) to generate bounding boxes.

2. Crops each bounding box from the original image.

3. Runs a classification (model2) on each cropped image.

4. Patches the original detection results with the classification labels.

Restriction: first model must be object detection, second model must be classification.

Functions

__init__(model1, ...)

__init__(model1, model2, crop_extent=0.0, crop_extent_option=CropExtentOptions.ASPECT_RATIO_NO_ADJUSTMENT)

Constructor.

Parameters:

Name	Type	Description	Default
model1	ModelLike	An object detection model producing bounding boxes.	required
model2	ModelLike	A classification model to classify each cropped region.	required
crop_extent	float	Extent of cropping (in percent of bbox size).	0.0
crop_extent_option	CropExtentOptions	Specifies how to adjust the bounding box before cropping.	ASPECT_RATIO_NO_ADJUSTMENT

predict_batch(data)

predict_batch(data)

Perform the full inference lifecycle for all objects in the given iterator (for example, list), but patch model1 bounding box labels with classification results from model2.

Parameters:

Name	Type	Description	Default
data	iterator	Iterator of input frames for model1. Each element returned by this iterator should be compatible with regular PySDK models.	required

Returns:

Type	Description
Iterator[InferenceResults]	Yields the detection results with patched classification labels after each frame completes.

transform_result2(result2)

transform_result2(result2)

Transform (patch) the classification result into the original detection results.

Parameters:

Name	Type	Description	Default
result2	InferenceResults	Classification result of model2.	required

Returns:

Type	Description
InferenceResults or None	The detection result (from model1) patched with classification labels, or None if we haven't moved to a new frame yet.

CroppingAndDetectingCompoundModel

CroppingAndDetectingCompoundModel

Bases: CroppingCompoundModel

Compound model class which

1. Uses an object detection model (model1) to generate bounding boxes (ROIs).

2. Crops each bounding box from the original image.

3. Uses another object detection model (model2) to further detect objects in each cropped region.

4. Combines the results of the second model from all cropped regions, mapping coords back to the original image.

Optionally, you can add model1 detections to the final result and/or apply NMS.

When model1 results are added, each detection from model2 will have a crop_index field, indicating which bounding box from model1 it corresponds to.

Restriction

First model should be object detection or pseudo-detection model like RegionExtractionPseudoModel, second model should be object detection.

Functions

__init__(model1, ...)

__init__(model1, model2, *, crop_extent=0.0, crop_extent_option=CropExtentOptions.ASPECT_RATIO_NO_ADJUSTMENT, add_model1_results=False, nms_options=None)

Constructor.

Parameters:

Name	Type	Description	Default
model1	ModelLike	Object detection model (or pseudo-detection).	required
model2	ModelLike	Object detection model.	required
crop_extent	float	Extent of cropping in percent of bbox size.	0.0
crop_extent_option	CropExtentOptions	Method of applying extended crop to the input image for model2.	ASPECT_RATIO_NO_ADJUSTMENT
add_model1_results	bool	If True, merges model1 detections into the final combined result. Each detection from model2 will have a crop_index field, indicating which bounding box from model1 it corresponds to.	False
nms_options	Optional[NmsOptions]	If provided, applies non-maximum suppression (NMS) to the combined result.	None

predict_batch(data)

predict_batch(data)

Perform the full inference lifecycle for all objects in the given iterator object (for example, list):

1. model1 detects or extracts bounding boxes (ROIs).

2. Each ROI is passed to model2 for detection.

3. model2 results for each ROI are merged and mapped back to original coordinates.

4. (Optional) NMS is applied and results from model1 can be included.

Parameters:

Name	Type	Description	Default
data	iterator	Iterator of input frames for model1. Each element returned by this iterator should be compatible with regular PySDK models.	required

Returns:

Type	Description
Iterator[InferenceResults]	Generator object which iterates over final detection results with possibly merged bounding boxes, adjusted to original image coordinates.

transform_result2(result2)

transform_result2(result2)

Combine detection results from model2 for each bbox from model1, translating coordinates back to the original image space.

Parameters:

Name	Type	Description	Default
result2	InferenceResults	Detection result of the second model.	required

Returns:

Type	Description
InferenceResults or None	The final detection results for the previous frame if a new frame started, or None otherwise.

RegionExtractionPseudoModel

RegionExtractionPseudoModel

Bases: ModelLike

Pseudo-model class which extracts regions from a given image according to given ROI boxes.

Attributes

custom_postprocessor

custom_postprocessor

property writable

Custom postprocessor class. Required for attaching analyzers to the pseudo-model.

When set, this replaces the default postprocessor with a user-defined postprocessor.

Returns:

Type	Description
Optional[type]	The user-defined postprocessor class, or None if not set.

non_blocking_batch_predict

non_blocking_batch_predict

property writable

Controls non-blocking mode for predict_batch().

Returns:

Type	Description
bool	True if non-blocking mode is enabled; otherwise False.

Functions

__getattr__(attr)

__getattr__(attr)

Fallback for getters of model-like attributes to model2.

__init__(roi_list, ...)

__init__(roi_list, model2, *, motion_detect=None)

Constructor.

Parameters:

Name	Type	Description	Default
roi_list	Union[list, ndarray]	Can be: - list of ROI boxes in [x1, y1, x2, y2] format, - 2D NumPy array of shape (N, 4), - 3D NumPy array of shape (K, M, 4), which will be flattened.	required
model2	Model	The second model in the pipeline.	required
motion_detect	Optional[MotionDetectOptions]	* When None, disabled motion detection. * When not None, applies motion detection before extracting ROI boxes. Boxes without motion are skipped.	None

predict_batch(data)

predict_batch(data)

Perform a pseudo-inference that outputs bounding boxes defined in roi_list.

If motion detection is enabled, skip ROIs where motion is not detected.

Parameters:

Name	Type	Description	Default
data	iterator	Iterator over the input images or frames. Each element returned by this iterator should be compatible with regular PySDK models.	required

Returns:

Type	Description
Iterator[InferenceResults or None]	Yields pseudo-inference results containing ROIs as bounding boxes, or yields None in non-blocking mode when no data is available.

NmsOptions

NmsOptions

dataclass

Options for non-maximum suppression (NMS) algorithm.

Attributes:

Name	Type	Description
threshold	float	IoU or IoS threshold for box clustering (range [0..1]).
use_iou	bool	If True, use IoU for box clustering, otherwise IoS.
box_select	NmsBoxSelectionPolicy	Box selection policy (e.g., keep the box with the highest probability).

MotionDetectOptions

MotionDetectOptions

dataclass

Options for motion detection algorithm.

Attributes:

Name	Type	Description
threshold	float	Threshold for motion detection [0..1], representing fraction of changed pixels relative to frame size.
look_back	int	Number of frames to look back to detect motion.