Class WorkflowProfiler

monai/utils/profiling.py:154–398

Profiler for timing all aspects of a workflow. This includes using stack tracing to capture call times for all selected calls (by default calls to `Transform.__call__` methods), times within context blocks, times to generate items from iterables, and times to execute decorated functions

class WorkflowProfiler:
    """
    Profiler for timing all aspects of a workflow. This includes using stack tracing to capture call times for
    all selected calls (by default calls to `Transform.__call__` methods), times within context blocks, times
    to generate items from iterables, and times to execute decorated functions.

    This profiler must be used only within its context because it uses an internal thread to read results from a
    multiprocessing queue. This allows the profiler to function across multiple threads and processes, though the
    multiprocess tracing is at times unreliable and not available in Windows at all.

    The profiler uses `sys.settrace` and `threading.settrace` to find all calls to profile, this will be set when
    the context enters and cleared when it exits so proper use of the context is essential to prevent excessive
    tracing. Note that tracing has a high overhead so times will not accurately reflect real world performance
    but give an idea of relative share of time spent.

    The tracing functionality uses a selector to choose which calls to trace, since tracing all calls induces
    infinite loops and would be terribly slow even if not. This selector is a callable accepting a `call` trace
    frame and returns True if the call should be traced. The default is `select_transform_call` which will return
    True for `Transform.__call__` calls only.

    Example showing use of all profiling functions:

    .. code-block:: python

        import monai.transform as mt
        from monai.utils import WorkflowProfiler
        import torch

        comp=mt.Compose([mt.ScaleIntensity(),mt.RandAxisFlip(0.5)])

        with WorkflowProfiler() as wp:
            for _ in wp.profile_iter("range",range(5)):
                with wp.profile_ctx("Loop"):
                    for i in range(10):
                        comp(torch.rand(1,16,16))

            @wp.profile_callable()
            def foo(): pass

            foo()
            foo()

        print(wp.get_times_summary_pd())  # print results

    Args:
        call_selector: selector to determine which calls to trace, use None to disable tracing
    """

    def __init__(self, call_selector=select_transform_call):
        self.results = defaultdict(list)
        self.parent_pid = os.getpid()
        self.read_thread: threading.Thread | None = None
        self.lock = threading.RLock()
        self.queue: multiprocessing.SimpleQueue = multiprocessing.SimpleQueue()
        self.queue_timeout = 0.1
        self.call_selector = call_selector

    def _is_parent(self):