Add num_workers to audonnx.Model and load()

audonnx/core/api.py

@@ -20,6 +20,7 @@ def load(
             typing.Tuple[str, typing.Dict],
             typing.Sequence[typing.Union[str, typing.Tuple[str, typing.Dict]]],
         ] = 'cpu',
+        num_workers: typing.Optional[int] = 1,
         auto_install: bool = False,
 ) -> Model:
     r"""Load model from folder.
@@ -44,6 +45,9 @@ def load(
         device: set device
             (``'cpu'``, ``'cuda'``, or ``'cuda:<id>'``)
             or a (list of) provider_
+        num_workers: number of threads for running
+            onnxruntime inference on cpu.
+            If ``None`` onnxruntime chooses the number of threads
         auto_install: install missing packages needed to create the object
 
     .. _provider: https://onnxruntime.ai/docs/execution-providers/
@@ -85,6 +89,7 @@ def load(
                 auto_install=auto_install,
                 override_args={
                     'device': device,
+                    'num_workers': num_workers,
                 },
             )
 
@@ -111,6 +116,7 @@ def load(
         labels=labels,
         transform=transform,
         device=device,
+        num_workers=num_workers,
     )
 
     return model

audonnx/core/model.py

@@ -50,6 +50,9 @@ class Model(audobject.Object):
         device: set device
             (``'cpu'``, ``'cuda'``, or ``'cuda:<id>'``)
             or a (list of) provider(s)_
+        num_workers: number of threads for running
+            onnxruntime inference on cpu.
+            If ``None`` onnxruntime chooses the number of threads
 
     Examples:
         >>> import audiofile
@@ -97,6 +100,7 @@ class Model(audobject.Object):
         },
         hide=[
             'device',
+            'num_workers',
         ],
     )
     def __init__(
@@ -106,6 +110,7 @@ def __init__(
             labels: Labels = None,
             transform: Transform = None,
             device: Device = 'cpu',
+            num_workers: typing.Optional[int] = 1,
     ):
         # keep original arguments to store them
         # when object is serialized
@@ -117,9 +122,15 @@ def __init__(
         self.path = audeer.path(path) if isinstance(path, str) else None
         r"""Model path"""
 
+        session_options = onnxruntime.SessionOptions()
+        if num_workers is not None:
+            session_options.inter_op_num_threads = num_workers
+            session_options.intra_op_num_threads = num_workers
+
         providers = device_to_providers(device)
         self.sess = onnxruntime.InferenceSession(
             self.path if isinstance(path, str) else path.SerializeToString(),
+            sess_options=session_options,
             providers=providers,
         )
         r"""Interference session"""

audonnx/core/testing.py

@@ -15,6 +15,7 @@ def create_model(
         dtype: int = onnx.TensorProto.FLOAT,
         opset_version: int = 14,
         device: Device = 'cpu',
+        num_workers: typing.Optional[int] = 1,
 ) -> Model:
     r"""Create test model.
 
@@ -36,6 +37,9 @@ def create_model(
         device: set device
             (``'cpu'``, ``'cuda'``, or ``'cuda:<id>'``)
             or a (list of) provider(s)_
+        num_workers: number of threads for running
+            onnxruntime inference on cpu.
+            If ``None`` onnxruntime chooses the number of threads
 
     Returns:
         model object
@@ -99,6 +103,7 @@ def create_model(
         object,
         transform=transform,
         device=device,
+        num_workers=num_workers,
     )
 
     return model

tests/test_model.py

@@ -228,6 +228,20 @@ def test_call_concat(model, outputs, expected):
             np.testing.assert_equal(y, expected)
 
 
+@pytest.mark.parametrize('device', ['cpu', 'cuda:0'])
+@pytest.mark.parametrize('num_workers', [None, 1, 2])
+def test_call_num_workers(device, num_workers):
+    model = audonnx.testing.create_model(
+        [[2]], device=device, num_workers=num_workers
+    )
+    y = model(
+        pytest.SIGNAL,
+        pytest.SAMPLING_RATE,
+    )
+    expected = np.array([0.0, 0.0], np.float32)
+    np.testing.assert_equal(y, expected)
+
+
 @pytest.mark.parametrize(
     'device',
     [