monai.deploy.operators.MonaiSegInferenceOperator

class monai.deploy.operators.MonaiSegInferenceOperator(fragment, *args, roi_size, pre_transforms, post_transforms, app_context, model_name='', overlap=0.25, sw_batch_size=4, inferer=sliding_window, model_path=PosixPath('/home/docs/checkouts/readthedocs.org/user_builds/monai-deploy-app-sdk/checkouts/stable/docs/source/model/model.ts'), **kwargs)

Bases: InferenceOperator

This segmentation operator uses MONAI transforms and performs Simple or Sliding Window Inference.

This operator performs pre-transforms on a input image, inference using a given model, and post-transforms. The segmentation image is saved as a named Image object in memory.

If specified in the post transforms, results may also be saved to disk.

This operator uses the MONAI inference utils functions for sliding window and simple inference, and thus input parameters need to be as expected by these functions.

Any additional sliding window arguments not explicitly defined in this operator can be passed via **kwargs for forwarding to ‘sliding_window_inference’.

Named Input:

image: Image object of the input image.

Named Output:

seg_image: Image object of the segmentation image. Not requiring a ready receiver.

Creates a instance of this class.

Parameters:

• fragment (Fragment) – An instance of the Application class which is derived from Fragment.

• roi_size (Sequence[int], int, optional) – The window size to execute “SLIDING_WINDOW” evaluation. Applicable for “SLIDING_WINDOW” only.

• pre_transforms (Compose) – MONAI Compose object used for pre-transforms.

• post_transforms (Compose) – MONAI Compose object used for post-transforms.

• app_context (AppContext) – Object holding the I/O and model paths, and potentially loaded models.

• model_name (str, optional) – Name of the model. Default to “” for single model app.

• overlap (Sequence[float], float) – The amount of overlap between scans along each spatial dimension. Defaults to 0.25. Applicable for “SLIDING_WINDOW” only.

• sw_batch_size (int) – The batch size to run window slices. Defaults to 4. Applicable for “SLIDING_WINDOW” only.

• inferer (InfererType, str) – The type of inferer to use, “SIMPLE” or “SLIDING_WINDOW”. Defaults to “SLIDING_WINDOW”.

• model_path (Path) – Path to the model file. Defaults to model/models.ts of current working dir.

• **kwargs – any other sliding window parameters to forward (e.g. mode, cval, etc.).

Methods

__init__(fragment, *args, roi_size, ...[, ...])	Creates a instance of this class.
add_arg(*args, **kwargs)	Overloaded function.
add_dynamic_flow(*args, **kwargs)	Overloaded function.
compute(op_input, op_output, context)	Infers with the input image and save the predicted image to output
compute_impl(input_image, context)
enable_metadata(self, enable)	Configure whether or not the metadata feature is enabled for this operator.
filter_sw_kwargs(**kwargs)	Filters the keyword arguments into a tuple of two dictionaries:
find_all_flow_info(self, predicate)	Find all flow info objects in the operator's next flows that match a given condition.
find_flow_info(self, predicate)	Find a flow info in the operator's next flows based on a given predicate.
initialize()	Default implementation of initialize
post_process(data, *args, **kwargs)	Transforms the prediction results from the model(s).
pre_process(data, *args, **kwargs)	Transforms input before being used for predicting on a model.
predict(data, *args, **kwargs)	Predicts results using the models(s) with input tensors.
queue_policy(self, port_name, port_type, policy)	Set the queue policy to be used by an input (or output) port's receiver (or transmitter).
receiver(self, port_name)	Get the receiver used by an input port.
resource(self, name)	Resources associated with the operator.
service(self, service_type[, id])	Retrieve a registered fragment service through the component's fragment.
setup(spec)	Default implementation of setup method.
start()	Default implementation of start
stop()	Default implementation of stop
stop_execution(self)	Stop the execution of the operator.
transmitter(self, port_name)	Get the transmitter used by an output port.

Attributes

INPUT_EXEC_PORT_NAME
MODEL_LOCAL_PATH
OUTPUT_EXEC_PORT_NAME
args	The list of arguments associated with the component.
async_condition	The internal asynchronous condition for the operator.
conditions	Conditions associated with the operator.
description	YAML formatted string describing the operator.
execution_context	The execution context for the operator.
fragment	The fragment (holoscan.core.Fragment) that the operator belongs to.
id	The identifier of the component.
inferer	The type of inferer to use
input_dataset_key	This is the input image key name used in dictionary based MONAI pre-transforms.
is_metadata_enabled	Boolean indicating whether the fragment this operator belongs to has metadata transmission enabled.
metadata	The metadata dictionary (holoscan.core.MetadataDictionary) associated with the operator.
metadata_policy	The metadata policy (holoscan.core.MetadataPolicy) associated with the operator.
name	The name of the operator.
next_flows	Get the list of flow information for connections to downstream operators.
operator_type	The operator type.
overlap	This is the overlap used during sliding window inference
pred_dataset_key	This is the prediction key name used in dictionary based MONAI post-transforms.
resources	Resources associated with the operator.
roi_size	The ROI size of tensors used in prediction.
spec	The operator spec (holoscan.core.OperatorSpec) associated with the operator.
sw_batch_size	The batch size to run window slices

class OperatorType(self: holoscan.core._core.OperatorBase.OperatorType, value: int)

Bases: pybind11_object

Enum class for operator types used by the executor.

• NATIVE: Native operator.

• GXF: GXF operator.

• VIRTUAL: Virtual operator. (for internal use, not intended for use by application authors)

Members:

NATIVE

GXF

VIRTUAL

property name

__init__(fragment, *args, roi_size, pre_transforms, post_transforms, app_context, model_name='', overlap=0.25, sw_batch_size=4, inferer=sliding_window, model_path=PosixPath('/home/docs/checkouts/readthedocs.org/user_builds/monai-deploy-app-sdk/checkouts/stable/docs/source/model/model.ts'), **kwargs)

Creates a instance of this class.

Parameters:

• fragment (Fragment) – An instance of the Application class which is derived from Fragment.

• roi_size (Sequence[int], int, optional) – The window size to execute “SLIDING_WINDOW” evaluation. Applicable for “SLIDING_WINDOW” only.

• pre_transforms (Compose) – MONAI Compose object used for pre-transforms.

• post_transforms (Compose) – MONAI Compose object used for post-transforms.

• app_context (AppContext) – Object holding the I/O and model paths, and potentially loaded models.

• model_name (str, optional) – Name of the model. Default to “” for single model app.

• overlap (Sequence[float], float) – The amount of overlap between scans along each spatial dimension. Defaults to 0.25. Applicable for “SLIDING_WINDOW” only.

• sw_batch_size (int) – The batch size to run window slices. Defaults to 4. Applicable for “SLIDING_WINDOW” only.

• inferer (InfererType, str) – The type of inferer to use, “SIMPLE” or “SLIDING_WINDOW”. Defaults to “SLIDING_WINDOW”.

• model_path (Path) – Path to the model file. Defaults to model/models.ts of current working dir.

• **kwargs – any other sliding window parameters to forward (e.g. mode, cval, etc.).

add_arg(*args, **kwargs)

Overloaded function.

1. add_arg(self: holoscan.core._core.OperatorBase, arg: holoscan.core._core.Arg) -> None

Add an argument to the component.

2. add_arg(self: holoscan.core._core.OperatorBase, arg: holoscan.core._core.ArgList) -> None

Add a list of arguments to the component.

3. add_arg(self: holoscan.core._core.OperatorBase, **kwargs) -> None

Add arguments to the component via Python kwargs.

4. add_arg(self: holoscan.core._core.OperatorBase, arg: holoscan.core._core.ConditionBase) -> None

5. add_arg(self: holoscan.core._core.OperatorBase, arg: holoscan.core._core.ResourceBase) -> None

Add a condition or resource to the Operator.

This can be used to add a condition or resource to an operator after it has already been constructed.

Parameters

argholoscan.core.Condition or holoscan.core.Resource

The condition or resource to add.

add_dynamic_flow(*args, **kwargs)

Overloaded function.

1. add_dynamic_flow(self: holoscan.core._core.OperatorBase, flow: holoscan.core._core.FlowInfo) -> None

2. add_dynamic_flow(self: holoscan.core._core.OperatorBase, flows: list[holoscan.core._core.FlowInfo]) -> None

3. add_dynamic_flow(self: holoscan.core._core.OperatorBase, next_op: holoscan.core._core.OperatorBase, next_input_port_name: str = ‘’) -> None

Add a dynamic flow from this operator to another operator.

Parameters

next_opholoscan.core.Operator

The downstream operator to connect to.

next_input_port_namestr, optional

The name of the input port on the downstream operator to connect to. If not specified, the first available input port will be used.

Notes

This method has several overloads to support different ways of creating dynamic flows:

1. add_dynamic_flow(next_op: Operator, next_input_port_name: str = ‘’) - Basic connection using default output port. This is the simplest form for connecting

   two operators when you only need to specify the destination.

2. add_dynamic_flow(curr_output_port_name: str, next_op: Operator, next_input_port_name: str = ‘’) - Connection with explicit output port specification. Use this when the source operator has

   multiple output ports and you need to specify which one to use.

3. add_dynamic_flow(flow: FlowInfo) - Connection using a FlowInfo object, which encapsulates all connection details including:

   • Source operator and its output port specification

   • Destination operator and its input port specification

   • Port names and associated IOSpecs

   • This is useful for complex connections or when reusing connection patterns.

4. add_dynamic_flow(flows: List[FlowInfo]) - Batch connection using multiple FlowInfo objects. Use this to set up multiple

   connections in a single call, which is more efficient than making multiple individual connections.

The FlowInfo class provides a complete description of a flow connection between operators, including all port specifications and naming. It’s particularly useful when you need to:

• Store and reuse connection patterns

• Create complex routing configurations

• Handle dynamic port specifications

• Manage multiple connections systematically

4. add_dynamic_flow(self: holoscan.core._core.OperatorBase, curr_output_port_name: str, next_op: holoscan.core._core.OperatorBase, next_input_port_name: str = ‘’) -> None

property args

The list of arguments associated with the component.

Returns

arglist : holoscan.core.ArgList

property async_condition

The internal asynchronous condition for the operator.

This property provides access to the internal asynchronous condition for the operator, which controls the scheduling of the operator’s compute method.

Returns

conditionholoscan.conditions.AsynchronousCondition

An instance of holoscan.conditions.AsynchronousCondition that is the internal asynchronous condition for the operator.

Notes

This object is only accessible after the executor has called Operator.initialize() via run() or run_async(). If accessed during Application.compose(), it will return None.

compute(op_input, op_output, context)

Infers with the input image and save the predicted image to output

Parameters:

• op_input (InputContext) – An input context for the operator.

• op_output (OutputContext) – An output context for the operator.

• context (ExecutionContext) – An execution context for the operator.

property conditions

Conditions associated with the operator.

property description

YAML formatted string describing the operator.

enable_metadata(self: holoscan.core._core.OperatorBase, enable: bool) → None

Configure whether or not the metadata feature is enabled for this operator. If it is not set, the default value will be determined by the enable_metadata setting from the Fragment that this operator belongs to.

property execution_context

The execution context for the operator.

This property provides access to the execution context of the operator, which contains information about the current execution environment including scheduling details.

Returns

holoscan.core.ExecutionContext

The execution context object for this operator.

static filter_sw_kwargs(**kwargs)

Filters the keyword arguments into a tuple of two dictionaries:

1. A dictionary of named parameters to pass to the sliding_window_inference function that:

   • Are not explicitly defined in the __init__ of this class

   • Are not explicitly used when calling sliding_window_inference

2. A dictionary of named parameters to pass to the base class __init__ of this class that:

   • Are not used when calling sliding_window_inference

   • Can be successfully converted from Python –> Holoscan’s C++ layer

Parameters:

**kwargs – extra arguments passed into __init__ beyond the explicitly defined args.

Returns:

A filtered dictionary of arguments to be passed to sliding_window_inference. filtered_base_init_params: A filtered dictionary of arguments to be passed to the base class __init__.

Return type:

filtered_swi_params

find_all_flow_info(self: holoscan.core._core.OperatorBase, predicate: Callable[[holoscan.core._core.FlowInfo], bool]) → list[holoscan.core._core.FlowInfo]

Find all flow info objects in the operator’s next flows that match a given condition.

Parameters

predicatecallable

A function that takes a FlowInfo object and returns a boolean.

Returns

list[holoscan.core.FlowInfo]

List of matching FlowInfo objects.

find_flow_info(self: holoscan.core._core.OperatorBase, predicate: Callable[[holoscan.core._core.FlowInfo], bool]) → holoscan.core._core.FlowInfo

Find a flow info in the operator’s next flows based on a given predicate.

Parameters

predicatecallable

A function that takes a FlowInfo object and returns a boolean.

Returns

holoscan.core.FlowInfo or None

The first matching FlowInfo object, or None if not found.

property fragment

The fragment (holoscan.core.Fragment) that the operator belongs to.

property id

The identifier of the component.

The identifier is initially set to -1, and will become a valid value when the component is initialized.

With the default executor (holoscan.gxf.GXFExecutor), the identifier is set to the GXF component ID.

Returns

id : int

property inferer: InfererType | str

The type of inferer to use

initialize()

Default implementation of initialize

property input_dataset_key

This is the input image key name used in dictionary based MONAI pre-transforms.

property is_metadata_enabled

Boolean indicating whether the fragment this operator belongs to has metadata transmission enabled.

property metadata

The metadata dictionary (holoscan.core.MetadataDictionary) associated with the operator.

property metadata_policy

The metadata policy (holoscan.core.MetadataPolicy) associated with the operator.

The supported policies are:

• MetadataPolicy.REJECT: Reject the new value if the key already exists

• MetadataPolicy.UPDATE: Replace existing value with the new one if the key already exists

• MetadataPolicy.INPLACE_UPDATE: Update the value stored within an existing MetadataObject in-place if the key already exists (in contrast to UPDATE which always replaces the existing MetadataObject with a new one).

• MetadataPolicy.RAISE: Raise an exception if the key already exists

property name

The name of the operator.

property next_flows

Get the list of flow information for connections to downstream operators.

Returns

list[holoscan.core.FlowInfo]

List of flow information objects describing connections to downstream operators.

property operator_type

The operator type.

holoscan.core.Operator.OperatorType enum representing the type of the operator. The two types currently implemented are native and GXF.

property overlap

This is the overlap used during sliding window inference

post_process(data, *args, **kwargs)

Transforms the prediction results from the model(s).

This method must be overridden by a derived class. Expected return is monai.transforms.Compose.

Parameters:

data (monai.transforms.Compose) – The pre-processing transforms in a Compose object.

Return type:

Union[Any, Image, Tuple[Any, ...], Dict[Any, Any]]

Returns:

monai.transforms.Compose encapsulating post-processing transforms.

Raises:

NotImplementedError – When the subclass does not override this method.

pre_process(data, *args, **kwargs)

Transforms input before being used for predicting on a model.

This method must be overridden by a derived class. Expected return is monai.transforms.Compose.

Parameters:

data (monai.data.ImageReader) – Reader used in LoadImage to load monai.deploy.core.Image as the input.

Return type:

Union[Any, Image, Tuple[Any, ...], Dict[Any, Any]]

Returns:

monai.transforms.Compose encapsulating pre transforms

Raises:

NotImplementedError – When the subclass does not override this method.

property pred_dataset_key

This is the prediction key name used in dictionary based MONAI post-transforms.

predict(data, *args, **kwargs)

Predicts results using the models(s) with input tensors.

This method is currently not used in this class, instead monai.inferers.sliding_window_inference is used.

Raises:

NotImplementedError – When the subclass does not override this method.

Return type:

Union[Image, Any, Tuple[Any, ...], Dict[Any, Any]]

queue_policy(self: holoscan.core._core.OperatorBase, port_name: str, port_type: holoscan.core._core.IOSpec.IOType = <IOType.INPUT: 0>, policy: holoscan.core._core.IOSpec.QueuePolicy = <QueuePolicy.FAULT: 2>) → None

Set the queue policy to be used by an input (or output) port’s receiver (or transmitter).

Parameters

port_namestr

The name of the port.

port_typeIOSpec.IOType, optional

Enum indicating whether port_name corresponds to an input port or output port.

policyIOSpec.QueuePolicy, optional

The queue policy to set. Valid values are:

• QueuePolicy.POP : If the queue is full, pop the oldest item, then add the new one.

• QueuePolicy.REJECT : If the queue is full, reject (discard) the new item.

• QueuePolicy.FAULT : If the queue is full, log a warning and reject the new item.

Returns

transmitterholoscan.resources.Transmitter or None

The transmitter used by this output port. Will be None if the port does not exist.

receiver(self: holoscan.core._core.OperatorBase, port_name: str) → Optional[holoscan::Receiver]

Get the receiver used by an input port.

Parameters

port_namestr

The name of the input port.

Returns

receiverholoscan.resources.Receiver

The receiver used by this input port. Will be None if the port does not exist.

resource(self: holoscan.core._core.OperatorBase, name: str) → object | None

Resources associated with the operator.

Parameters

namestr

The name of the resource to retrieve

Returns

holoscan.core.Resource or None

The resource with the given name. If no resource with the given name is found, None is returned.

property resources

Resources associated with the operator.

property roi_size

The ROI size of tensors used in prediction.

service(self: object, service_type: type, id: str = '') → object

Retrieve a registered fragment service through the component’s fragment.

This method delegates to the fragment’s service() method to retrieve a previously registered fragment service by its type and optional identifier. Returns None if no fragment service is found with the specified type and identifier.

Parameters

service_typetype

The type of the fragment service to retrieve. Must be a type that inherits from Resource or FragmentService.

idstr, optional

The identifier of the fragment service. If empty, retrieves by service type only. For Resources, this would typically be the resource’s name.

Returns

object or None

The fragment service instance of the requested type, or None if not found. If the service wraps a Resource and a Resource type is requested, the unwrapped Resource instance is returned.

Raises

RuntimeError

If the component has no associated fragment or if the fragment’s service method cannot be accessed.

Notes

This is a convenience method that internally calls the fragment’s service() method. For services that wrap Resources, the method will automatically unwrap and return the Resource if a Resource type is requested.

setup(spec)

Default implementation of setup method.

property spec

The operator spec (holoscan.core.OperatorSpec) associated with the operator.

start()

Default implementation of start

stop()

Default implementation of stop

stop_execution(self: holoscan.core._core.OperatorBase) → None

Stop the execution of the operator.

This method is used to stop the execution of the operator by setting the internal async condition to EVENT_NEVER state, which sets the scheduling condition to NEVER. Once stopped, the operator will not be scheduled for execution (the compute() method will not be called).

Note that executing this method does not trigger the operator’s stop() method. The stop() method is called only when the scheduler deactivates all operators together.

property sw_batch_size

The batch size to run window slices

transmitter(self: holoscan.core._core.OperatorBase, port_name: str) → Optional[holoscan::Transmitter]

Get the transmitter used by an output port.

Parameters

port_namestr

The name of the output port.

Returns

transmitterholoscan.resources.Transmitter or None

The transmitter used by this output port. Will be None if the port does not exist.