holoscan-cu13 3.7.0__cp310-cp310-manylinux_2_35_aarch64.whl

holoscan_cu13-3.7.0.data/purelib/holoscan/decorator.py

@@ -0,0 +1,662 @@
+"""
+SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+SPDX-License-Identifier: Apache-2.0
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""  # noqa: E501
+
+"""This module provides a decorator API for creating Python Operators
+
+.. autosummary::
+
+    holoscan.decorator.create_op
+    holoscan.decorator.Input
+    holoscan.decorator.Output
+"""
+
+import ast
+import inspect
+import textwrap
+from collections.abc import Callable, Sequence
+from dataclasses import dataclass, field
+from typing import Any
+
+import cupy as cp
+import numpy as np
+
+from holoscan.conditions import BooleanCondition
+from holoscan.core import (
+    ConditionType,
+    IOSpec,
+    Operator,
+    OperatorSpec,
+)
+from holoscan.core import _Condition as ConditionBase
+from holoscan.core import _Resource as ResourceBase
+from holoscan.core._core import Fragment as FragmentBase
+from holoscan.core._core import Tensor as TensorBase
+
+__all__ = ["Input", "Output", "create_op"]
+
+
+def _is_tensor_like(obj):
+    return (
+        (hasattr(obj, "__dlpack__") and hasattr(obj, "__dlpack_device__"))
+        or hasattr(obj, "__cuda_array_interface__")
+        or hasattr(obj, "__array_interface__")
+    )
+
+
+def _as_python_tensor(tensor):
+    if hasattr(tensor, "__array_interface__") or (
+        hasattr(tensor, "__dlpack_device__") and tensor.__dlpack_device__()[0] == 1
+    ):
+        return np.asarray(tensor)
+    else:
+        return cp.asarray(tensor)
+
+
+@dataclass
+class Input:
+    """Class for specifying an input port and how the received value maps to a function's arguments.
+
+    Parameters
+    ----------
+    name : str
+        The name of the input port.
+    arg_map: str or dict[str, str]
+        If `arg_map` is a str, the Python object received by the input port is passed to the
+        function argument specified by `arg_map`. If `arg_map` is a dict, the input is assumed to be
+        a TensorMap (dictionary of tensors). In this case the keys of the dict are the tensor names
+        and the values are the names of the function arguments that the tensors map to.
+    size: int | holoscan.core.IOSpec.IOSize, optional
+        The size of the queue for the input port.
+        By default, `IOSpec.SIZE_ONE` (== `IOSpec.IOSize(1)`) is used.
+        If `IOSpec.ANY_SIZE` is used, it defines multiple receivers internally for the input port.
+        Otherwise, the size of the input queue is set to the specified value, and the message
+        available condition for the input port is set with `min_size` equal to the same value.
+
+        The following size constants are supported:
+        - ``IOSpec.ANY_SIZE``: Any size.
+        - ``IOSpec.PRECEDING_COUNT``: Number of preceding connections.
+        - ``IOSpec.SIZE_ONE``: The queue size is 1.
+
+        Please refer to the [Holoscan SDK User Guide](https://docs.nvidia.com/holoscan/sdk-user-guide/holoscan_create_operator.html#receiving-any-number-of-inputs-python)
+        to see how to receive any number of inputs in Python.
+    policy : holoscan.core.IOSpec.QueuePolicy, optional
+        The queue policy for the input port.
+        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.
+    condition_type : holoscan.core.ConditionType, optional
+        The condition type for the input port.
+    condition_kwargs : dict[str, Any], optional
+        The keywords passed onto the condition specified by `condition_type`.
+    connector_type : holoscan.core.IOSpec.ConnectorType, optional
+        The connector type for the input port.
+    connector_kwargs : dict[str, Any], optional
+        The keywords passed onto the connector specified by `connector_type`.
+    """
+
+    name: str
+    arg_map: str | dict[str, str] | None = ()
+    size: int | IOSpec.IOSize = IOSpec.SIZE_ONE
+    policy: IOSpec.QueuePolicy | None = None
+    condition_type: ConditionType | None = None
+    condition_kwargs: dict[str, Any] = field(default_factory=dict)
+    connector_type: IOSpec.ConnectorType | None = None
+    connector_kwargs: dict[str, Any] = field(default_factory=dict)
+
+    def create_input(self, spec: OperatorSpec) -> IOSpec:
+        if isinstance(self.size, int):
+            self.size = IOSpec.IOSize(self.size)
+        elif not isinstance(self.size, IOSpec.IOSize):
+            raise ValueError(f"Invalid size: {self.size}")
+
+        if self.policy is not None and not isinstance(self.policy, IOSpec.QueuePolicy):
+            raise ValueError(f"Invalid policy: {self.policy}")
+
+        iospec = spec.input(self.name, size=self.size, policy=self.policy)
+
+        if self.condition_type is not None:
+            iospec = iospec.condition(self.condition_type, **self.condition_kwargs)
+        if self.connector_type is not None:
+            iospec = iospec.connector(self.connector_type, **self.connector_kwargs)
+
+
+@dataclass
+class Output:
+    """Class for specifying an output port and how one or more of a functions returned value(s) map
+    to it.
+
+    Parameters
+    ----------
+    name : str
+        The name of the output port.
+    tensor_names: str, tuple(str) or None
+        If None, whatever Python object the func outputs is emitted on the output port. If a tuple
+        of strings is provided it is assumed that the func returns a dictionary of tensors. The
+        names in the tuple specify which tensors in the dict will be transmitted on the output
+        port. There is no need to specify `tensor_names` if all tensors in a dict returned by the
+        function are to be transmitted. In the case of a single tensor name, a string can be
+        provided instead of a tuple.
+    size: int | holoscan.core.IOSpec.IOSize, optional
+        The size of the queue for the output port.
+        By default, `IOSpec.SIZE_ONE` (== `IOSpec.IOSize(1)`) is used.
+    policy : holoscan.core.IOSpec.QueuePolicy, optional
+        The queue policy for the output port.
+        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.
+    condition_type : holoscan.core.ConditionType, optional
+        The condition type for the input port.
+    condition_kwargs : dict[str, Any], optional
+        The keywords passed onto the condition specified by `condition_type`.
+    connector_type : holoscan.core.IOSpec.ConnectorType, optional
+        The connector type for the input port.
+    connector_kwargs : dict[str, Any], optional
+        The keywords passed onto the connector specified by `connector_type`.
+    """
+
+    name: str
+    tensor_names: str | tuple[str] | None = ()
+    size: int | IOSpec.IOSize = IOSpec.SIZE_ONE
+    policy: IOSpec.QueuePolicy | None = None
+    condition_type: ConditionType | None = None
+    condition_kwargs: dict[str, Any] = field(default_factory=dict)
+    connector_type: IOSpec.ConnectorType | None = None
+    connector_kwargs: dict[str, Any] = field(default_factory=dict)
+
+    def create_output(self, spec: OperatorSpec) -> IOSpec:
+        if isinstance(self.size, int):
+            self.size = IOSpec.IOSize(self.size)
+        elif not isinstance(self.size, IOSpec.IOSize):
+            raise ValueError(f"Invalid size: {self.size}")
+
+        if self.policy is not None and not isinstance(self.policy, IOSpec.QueuePolicy):
+            raise ValueError(f"Invalid policy: {self.policy}")
+
+        iospec = spec.output(self.name, size=self.size, policy=self.policy)
+        if self.condition_type is not None:
+            iospec = iospec.condition(self.condition_type, **self.condition_kwargs)
+        if self.connector_type is not None:
+            iospec = iospec.connector(self.connector_type, **self.connector_kwargs)
+
+
+def _as_input(input_: str | Input):
+    """Cast str to Output object."""
+    if isinstance(input_, str):
+        return Input(input_, arg_map=input_)
+    elif not isinstance(input_, Input):
+        return ValueError("`inputs` must be a single port name or Input object or a tuple of these")
+    return input_
+
+
+def _as_output(output: str | Output):
+    """Cast str to Output object."""
+    if isinstance(output, str):
+        return Output(output)
+    elif not isinstance(output, Output):
+        return ValueError(
+            "`outputs` must be a single port name or Output object or a tuple of these"
+        )
+    return output
+
+
+def _has_function_returns_value(func):
+    """Check if the provided function has any return statements returning a value."""
+
+    class ReturnVisitor(ast.NodeVisitor):
+        def __init__(self):
+            self.returns_value = False
+
+        def visit_Return(self, node):  # noqa: N802
+            # check if the return statement has a value
+            if node.value is not None:
+                self.returns_value = True
+                return
+
+            self.generic_visit(node)
+
+        def visit_ClassDef(self, node):  # noqa: N802, ARG002
+            return
+
+        def visit_FunctionDef(self, node):  # noqa: N802, ARG002
+            return
+
+        def visit_AsyncFunctionDef(self, node):  # noqa: N802, ARG002
+            return
+
+        def visit(self, node):
+            if self.returns_value:
+                return
+            super().visit(node)
+
+    # parse the source code into an AST
+    source_code = inspect.getsource(func)
+    # deindent the text if it is indented
+    source_code = textwrap.dedent(source_code)
+    tree = ast.parse(source_code)
+    # initialize the visitor
+    visitor = ReturnVisitor()
+    # walk the AST
+    for node in ast.walk(tree):
+        if isinstance(node, ast.FunctionDef) and node.name == func.__name__:
+            visitor.generic_visit(node)
+            break
+    return visitor.returns_value
+
+
+def create_op(
+    function_or_class: type | Callable[..., Any] | None = None,
+    inputs: str | Input | Sequence[str | Input] = (),
+    outputs: str | Output | Sequence[str | Output] = (),
+    cast_tensors: bool = True,
+    op_param: str | None = None,
+) -> Callable:
+    """Decorator for creating an operator from a function or a class.
+
+    When the decorator is used on a class, the class must have a `__call__` method that will be
+    used as the operator function.
+
+    inputs : str, Input, or Tuple[str | Input], optional
+        If a str is provided, it is assumed to be the name of the input port and that the function
+        has a variable matching that port name to which the object received on the port will be
+        connected. If the port name does not match the name of the variable in the function
+        signature, or if there are multiple tensors to be mapped to multiple objects, use an Input
+        argument. A tuple of str or Input objects can be provided to specify multiple input ports.
+        The default of an empty tuple corresponds to no input ports.
+    outputs : str, Output, or Tuple[str | Output], optional
+        If a str is provided, any value returned by the function will be emitted on an output port
+        of that name. If a tuple of multiple str is provided and the function returns a tuple, then
+        the tuple elements will be emitted from each output port in the order at which they are
+        defined. In this case, the number of output ports should match the length of the output
+        tuple. Finally, an Output object can be provided in the case that the function returns a
+        dictionary of output arrays that should be split across multiple ports.
+    cast_tensors : bool, optional
+        If True, automatically cast any tensor-like input to a NumPy or CuPy array (for host and
+        device tensors, respectively). If set to False, these will be left as `holoscan.Tensor` and
+        the user will have to cast to the desired type within the body of the decorated function or
+        class.
+    op_param : str, optional
+        If provided, adds this parameter name to the function signature which will
+        contain a reference to the operator instance. This allows the function to
+        access operator methods and attributes.
+
+    Notes
+    -----
+    Another case where using `Input` or `Output` objects is necessary is if the user wishes to
+    override the default connector or condition types for the port.
+    """
+    # used to store the class object if the decorator is used on a class
+    class_obj = None
+    # used to determine if the decorator was used without args
+    is_without_args = function_or_class is not None
+
+    # convert scalars to tuple
+    if isinstance(inputs, str | Input):
+        inputs = (inputs,)
+    # convert any str in the tuple to an Input object
+    inputs = tuple(_as_input(i) for i in inputs)
+
+    if isinstance(outputs, str | Output):
+        outputs = (outputs,)
+    # convert any str in the tuple to an Output object
+    outputs = tuple(_as_output(o) for o in outputs)
+
+    if not isinstance(outputs, tuple):
+        raise ValueError(
+            "`outputs` must be a single port name or Output object or a tuple of these"
+        )
+
+    if op_param is not None and not isinstance(op_param, str):
+        raise TypeError(f"op_param must be a string or None, got {type(op_param)}")
+
+    def decorator(func_or_cls):
+        nonlocal function_or_class, class_obj
+
+        def make_class(*args, **kwargs):
+            if "fragment" in kwargs:
+                fragment = kwargs.pop("fragment")
+            elif args and isinstance(args[0], FragmentBase):
+                fragment, args = args[0], args[1:]
+            else:
+                raise ValueError(
+                    "fragment must be provided via kwarg or as the first positional argument"
+                )
+
+            # frame = inspect.currentframe()
+            # args_names, _, _, locals_dict = inspect.getargvalues(frame)
+            # print(f"{args_names=}, {locals_dict=}")
+
+            class DynamicOp(Operator):
+                def __init__(
+                    self,
+                    fragment: FragmentBase,
+                    *args,
+                    inputs,
+                    outputs,
+                    cast_tensors=cast_tensors,
+                    op_param=op_param,
+                    **kwargs,
+                ):
+                    self.func = func_or_cls
+                    self.input_objs = inputs
+                    self.output_objs = outputs
+                    self.is_generator = inspect.isgeneratorfunction(self.func)
+                    self.gen_obj = None
+                    self.cast_tensors = cast_tensors
+                    self.op_param = op_param
+
+                    # remove conditions and resources from *args
+                    condition_args = tuple(a for a in args if isinstance(a, ConditionBase))
+                    resource_args = tuple(a for a in args if isinstance(a, ResourceBase))
+                    args = tuple(a for a in args if not isinstance(a, ConditionBase | ResourceBase))
+                    self.func_args = args
+
+                    # add a boolean condition to prevent triggering if the function is a generator
+                    # and the iteration is complete
+                    if self.is_generator:
+                        condition_args = condition_args + (
+                            BooleanCondition(fragment, name="_generator_func"),
+                        )
+
+                    # set name kwarg to self.func.__name__ if not provided
+                    name = kwargs.pop("name", self.func.__name__)
+
+                    argspec = inspect.getfullargspec(self.func)
+
+                    # remove self from argspec.args if the decorator is used on a class
+                    if class_obj:
+                        argspec = argspec._replace(args=argspec.args[1:])
+                        self.class_obj = class_obj
+
+                    self.func_argspec = argspec
+
+                    # populate inputs and outputs with defaults if decorator was used without args
+                    if is_without_args:
+                        self.input_objs = tuple(Input(name, arg_map=name) for name in argspec.args)
+                        # configure the output port if the function contains return statements
+                        # (in this case, the port name will be left empty)
+                        if _has_function_returns_value(function_or_class):
+                            self.output_objs = tuple((Output(""),))
+
+                    # populate all arguments not provided with defaults
+                    if argspec.kwonlydefaults is not None:
+                        for k in argspec.kwonlyargs:
+                            if k not in kwargs and k in argspec.kwonlydefaults:
+                                kwargs[k] = argspec.kwonlydefaults[k]
+
+                    # store a list of what ports map to what function arguments
+                    self.input_mappings = {}
+                    for input_obj in self.input_objs:
+                        # store what argument(s) this input maps to
+                        self.input_mappings[input_obj.name] = input_obj.arg_map
+
+                    # sets self.dynamic_kwargs and self.fixed_kwargs
+                    self._set_fixed_and_dynamic_kwargs(kwargs)
+
+                    # get the type annotations dict for the function (not currently used)
+                    # self.func_annotations = inspect.get_annotations(self.func)
+                    self.func_annotations = self.func.__annotations__
+
+                    super().__init__(fragment, *condition_args, *resource_args, name=name)
+
+                def _set_fixed_and_dynamic_kwargs(self, kwargs):
+                    """Split provided kwargs into those which are "fixed" and those which are
+                    "dynamic".
+
+                    Here "dynamic" refers to function arguments that are obtained from input
+                    ports. The keys for self.dynamic_kwargs are determined here, but the values
+                    are initialized to None. Actual values get set during each `compute` call.
+
+                    "fixed" refers to other keyword arguments to the function that don't change
+                    across calls.
+                    """
+                    self.dynamic_kwargs = {}
+                    for input_map in self.input_mappings.values():
+                        if isinstance(input_map, str):
+                            self._add_dynamic_arg(input_map, kwargs)
+                        elif isinstance(input_map, dict):
+                            for arg_name in input_map.values():
+                                self._add_dynamic_arg(arg_name, kwargs)
+                    self.fixed_kwargs = kwargs
+
+                    # add the operator instance to the kwargs if op_param was specified
+                    if self.op_param:
+                        self.fixed_kwargs[self.op_param] = self
+
+                    # store any positional args with specified defaults in fixed_kwargs instead
+                    argspec = self.func_argspec
+                    if argspec.defaults is not None:
+                        n_default_positional = len(argspec.defaults)
+                        if n_default_positional > 0:
+                            self.func_args = self.func_args[:-n_default_positional]
+                        n_required_positional = len(argspec.args) - len(argspec.defaults)
+                        for k, v in zip(
+                            argspec.args[n_required_positional:], argspec.defaults, strict=False
+                        ):
+                            # don't overwrite any kwargs that were provided
+                            if k not in self.fixed_kwargs:
+                                self.fixed_kwargs[k] = v
+
+                    # Now that all args with defaults are in self.fixed_kwargs we can check if any
+                    # of the required arguments were not specified
+                    required_args = set(argspec.args) | set(argspec.kwonlyargs)
+                    if argspec.kwonlydefaults is not None:
+                        required_args -= set(argspec.kwonlydefaults.keys())
+                    for arg in required_args:
+                        if arg not in self.fixed_kwargs and arg not in self.dynamic_kwargs:
+                            raise ValueError(f"required argument, '{arg}', has not been specified")
+
+                def _add_dynamic_arg(self, arg_name, kwargs):
+                    """helper function for _set_fixed_and_dynamic_kwargs"""
+                    if arg_name in self.dynamic_kwargs:
+                        raise ValueError(
+                            "duplicate specification of mapping to function kwarg: '{arg_name}'"
+                        )
+                    self.dynamic_kwargs[arg_name] = None
+                    try:
+                        kwargs.pop(arg_name)
+                    except KeyError as e:
+                        argspec = self.func_argspec
+                        if arg_name not in argspec.kwonlyargs + argspec.args:
+                            msg = (
+                                f"Provided func does not have an arg or kwarg named '{arg_name}'."
+                                " The provided wrapped function has"
+                                f" positional args: {argspec.args}"
+                                f" and keyword-only args: {argspec.kwonlyargs}"
+                            )
+                            raise KeyError(msg) from e
+
+                # # not used by the Application, but can be useful to test the call
+                # def __call__(self, *args, **kwargs):
+                #     print(f"{self.msg=}")
+                #     return self.func(*self.func_args, *args, **self.fixed_kwargs, **kwargs)
+
+                def setup(self, spec: OperatorSpec):
+                    for input_obj in self.input_objs:
+                        input_obj.create_input(spec)
+
+                    self.output_tensor_map = {}
+                    for output_obj in self.output_objs:
+                        output_obj.create_output(spec)
+                        if isinstance(output_obj.tensor_names, str):
+                            output_obj.tensor_names = (output_obj.tensor_names,)
+                        self.output_tensor_map[output_obj.name] = tuple(output_obj.tensor_names)
+
+                def compute(self, op_input, op_output, context):
+                    for port_name, arg_map in self.input_mappings.items():
+                        # print(f"input {port_name=}, {arg_map=}")
+                        msg = op_input.receive(port_name)
+                        if isinstance(arg_map, str):
+                            # print(f"{msg=}")
+                            if isinstance(msg, dict):
+                                try:
+                                    # try tensor based on matching name
+                                    msg = msg[arg_map]
+                                except KeyError as e:
+                                    # use tensor regardless of name if only one is present
+                                    tensors = tuple(
+                                        v for k, v in msg.items() if isinstance(v, TensorBase)
+                                    )
+                                    if len(tensors) == 1:
+                                        msg = tensors[0]
+                                    elif len(tensors) > 1:
+                                        raise ValueError(
+                                            "More than one tensor found in port, but none has "
+                                            f"name {arg_map}"
+                                        ) from e
+
+                            # cast holoscan.Tensor to cp.asarray(Tensor) here or require the user
+                            # to do it in the provided func?
+                            if self.cast_tensors and isinstance(msg, TensorBase):
+                                msg = _as_python_tensor(msg)
+
+                            self.dynamic_kwargs[arg_map] = msg
+                        elif isinstance(arg_map, dict):
+                            for tensor_name, arg_name in arg_map.items():
+                                try:
+                                    val = msg[tensor_name]
+                                except KeyError as e:
+                                    raise KeyError(
+                                        f"key with name '{tensor_name}' not found in input dict"
+                                    ) from e
+                                if self.cast_tensors and isinstance(val, TensorBase):
+                                    val = _as_python_tensor(val)
+                                self.dynamic_kwargs[arg_name] = val
+
+                    if self.is_generator:
+                        if self.gen_obj is None:
+                            out = self.func(
+                                *self.func_args, **self.fixed_kwargs, **self.dynamic_kwargs
+                            )
+                            self.gen_obj = out
+                        try:
+                            out = next(self.gen_obj)
+                        except StopIteration:
+                            # disable the condition to prevent further calls
+                            self.conditions["_generator_func"].disable_tick()
+                            return
+                    else:
+                        out = self.func(*self.func_args, **self.fixed_kwargs, **self.dynamic_kwargs)
+
+                    # if the output is a tuple and there is >1 port, we distribute the outputs
+                    if isinstance(out, tuple) and (len(self.output_tensor_map) > 1):
+                        # for tuple case, each port should correspond to each output tuple element
+                        if any([len(names) > 1 for names in self.output_tensor_map.values()]):
+                            raise ValueError(
+                                "The function output was found to be a tuple type, but each "
+                                "output tuple element must have its own port. In other words, "
+                                "the `outputs` argument of `create_op` should be a tuple of port "
+                                "names equal in length to the returned tuple."
+                            )
+                        # Make sure check that the output tuple length and number of ports match
+                        if len(out) != len(self.output_tensor_map):
+                            raise ValueError(
+                                f"The number of output tuple elements and number of tensors must "
+                                f"match.\n"
+                                f"Output tuple length = {len(out)}\n"
+                                f"Number of output tensors = {len(self.output_tensor_map)}"
+                            )
+                        for (port_name, tensor_names), out_element in zip(
+                            self.output_tensor_map.items(), out, strict=False
+                        ):
+                            if _is_tensor_like(out_element):
+                                name = "" if len(tensor_names) == 0 else tensor_names[0]
+                                out_element = {name: out_element}
+                            op_output.emit(out_element, port_name)
+                        return
+
+                    for port_name, tensor_names in self.output_tensor_map.items():
+                        if tensor_names is None or len(tensor_names) == 0:
+                            if _is_tensor_like(out):
+                                # emit as dict of tensor-like objects
+                                out = {"": out}
+                            op_output.emit(out, port_name)
+                        elif len(tensor_names) == 1:
+                            name = tensor_names[0]
+                            if _is_tensor_like(out):
+                                # emit as dict of tensor-like objects
+                                out = {name: out}
+                                op_output.emit(out, port_name)
+                            else:
+                                if name not in out:
+                                    raise ValueError(
+                                        f"tensor with name '{name}' not found in function output"
+                                    )
+                                op_output.emit({name: out[name]}, port_name)
+                        else:
+                            out_tensors = {}
+                            for name in tensor_names:
+                                if name not in out:
+                                    raise ValueError(
+                                        f"tensor with name '{name}' not found in function output"
+                                    )
+                                out_tensors[name] = out[name]
+                            # print(f"outputting tensors named: {tuple(out_tensors.keys())} on
+                            # port {port_name}")
+                            # print(f"tensormap emit of {out_tensors=}")
+                            op_output.emit(out_tensors, port_name)
+
+            op = DynamicOp(
+                fragment, *args, inputs=inputs, outputs=outputs, op_param=op_param, **kwargs
+            )
+
+            def _to_camel_case(name):
+                """Convert name to camel case"""
+                parts = name.split("_")
+                return "".join(p.capitalize() for p in parts)
+
+            # manually update instead of using functools.update_wrapper(op, func_or_cls) because:
+            #    - don't want to overwrite __doc__ with func.__doc__
+            #    - want to use name instead of func.__name__
+            if class_obj:
+                class_name = class_obj.__class__.__name__
+                op.__name__ = class_name + "Op" if not class_name.endswith("Op") else class_name
+            else:
+                op.__name__ = _to_camel_case(func_or_cls.__name__) + "Op"
+            op.__qualname__ = op.__name__
+            op.__module__ = func_or_cls.__module__
+            return op
+
+        def init_class(*args, **kwargs):
+            nonlocal class_obj, function_or_class
+            # create an instance of the class (using function_or_class as the class)
+            class_obj = function_or_class(*args, **kwargs)
+            # use the class's __call__ method as the operator function
+            if not callable(class_obj):
+                raise ValueError(
+                    f"{function_or_class} must have a __call__ method to be used as an operator"
+                )
+            function_or_class = class_obj.__call__
+            return decorator(function_or_class)
+
+        if func_or_cls is None:
+            return decorator
+
+        # check if the decorator was used on a class first
+        if inspect.isclass(func_or_cls):  # if isinstance(func_or_cls, type):
+            function_or_class = func_or_cls
+            return init_class
+
+        if callable(func_or_cls):
+            return make_class
+
+        raise Exception(f"Invalid usage of decorator for {func_or_cls}")
+
+    return decorator(function_or_class)

holoscan_cu13-3.7.0.data/purelib/holoscan/executors/__init__.py

@@ -0,0 +1,26 @@
+# SPDX-FileCopyrightText: Copyright (c) 2022-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""This module provides a Python API for the C++ API Executor classes.
+
+.. autosummary::
+
+    holoscan.executors.GXFExecutor
+"""
+
+from ._executors import GXFExecutor
+
+__all__ = [
+    "GXFExecutor",
+]