Serialize model (#37)

* Serialize model

* Add tests/python/precommit/test_serialize.py

* isort style

* Replace [] with ''

* Reuse model.parameters()

* Reuse tmp_path

* serialize->save

* black style

* ssd_mobilenet_fpn.serialize->ssd_mobilenet_fpn.save

Author: Wovchena

.gitignore

@@ -1,6 +1,10 @@
-# mask.png is generated by examples/python/simple_local/run.py
+# Generated by examples/python/synchronous_api/run.py
 mask.png
 ssd_mobilenet_v1_fpn_coco_with_preprocessing.xml
+ssd_mobilenet_v1_fpn_coco_with_preprocessing.bin
+tmp/
+# Generated by tests/python/accuracy/test_accuracy.py
+test_scope.json
 
 # Byte-compiled / optimized / DLL files
 __pycache__/

examples/python/synchronous_api/run.py

@@ -50,10 +50,7 @@ def main():
     )
     detections = ssd_mobilenet_fpn(image)
     print(f"Detection results: {detections}")
-    ov.serialize(
-        ssd_mobilenet_fpn.get_model(),
-        "ssd_mobilenet_v1_fpn_coco_with_preprocessing.xml",
-    )
+    ssd_mobilenet_fpn.save("ssd_mobilenet_v1_fpn_coco_with_preprocessing.xml")
 
     # Instantiate from a local model (downloaded previously)
     ssd_mobilenet_fpn_local = DetectionModel.create_model(

model_api/python/openvino/model_api/adapters/openvino_adapter.py

@@ -20,7 +20,7 @@
 
 try:
     import openvino.runtime as ov
-    from openvino.preprocess import ColorFormat, PrePostProcessor, ResizeAlgorithm
+    from openvino.preprocess import ColorFormat, PrePostProcessor
     from openvino.runtime import (
         AsyncInferQueue,
         Core,
@@ -131,7 +131,7 @@ def __init__(
         self,
         core,
         model,
-        weights_path=None,
+        weights_path="",
         model_parameters={},
         device="CPU",
         plugin_config=None,
@@ -170,8 +170,7 @@ def __init__(
                     "from buffer" if self.model_from_buffer else self.model_path
                 )
             )
-            weights = weights_path if self.model_from_buffer else ""
-            self.model = core.read_model(self.model_path, weights)
+            self.model = core.read_model(self.model_path, weights_path)
             return
         if isinstance(model, str):
             from openvino.model_zoo.models import OMZModel, list_models
@@ -409,10 +408,10 @@ def embed_preprocessing(
         self.load_model()
 
     def get_model(self):
-        """Returns the ov.Model object
+        """Returns the openvino.runtime.Model object
 
         Returns:
-            ov.Model object
+            openvino.runtime.Model object
         """
         return self.model
 

model_api/python/openvino/model_api/adapters/utils.py

@@ -19,7 +19,7 @@
 
 import numpy as np
 import openvino.runtime as ov
-from openvino.runtime import Output, layout_helpers
+from openvino.runtime import Output, Type, layout_helpers
 from openvino.runtime import opset10 as opset
 from openvino.runtime.utils.decorators import custom_preprocess_function
 
@@ -101,8 +101,8 @@ def resize_image_letterbox_graph(input: Output, size, interpolation="linear"):
         opset.gather(image_shape, opset.constant(h_axis), axis=0),
         destination_type="f32",
     )
-    w_ratio = opset.divide(opset.constant(w, dtype=float), iw)
-    h_ratio = opset.divide(opset.constant(h, dtype=float), ih)
+    w_ratio = opset.divide(opset.constant(w, dtype=Type.f32), iw)
+    h_ratio = opset.divide(opset.constant(h, dtype=Type.f32), ih)
     scale = opset.minimum(w_ratio, h_ratio)
     nw = opset.convert(opset.multiply(iw, scale), destination_type="i32")
     nh = opset.convert(opset.multiply(ih, scale), destination_type="i32")

model_api/python/openvino/model_api/models/image_model.py

@@ -89,11 +89,11 @@ def parameters(cls):
         parameters.update(
             {
                 "mean_values": ListValue(
-                    default_value=None,
+                    default_value=[],
                     description="Normalization values, which will be subtracted from image channels for image-input layer during preprocessing",
                 ),
                 "scale_values": ListValue(
-                    default_value=None,
+                    default_value=[],
                     description="Normalization values, which will divide the image channels for image-input layer",
                 ),
                 "reverse_input_channels": BooleanValue(

model_api/python/openvino/model_api/models/model.py

@@ -124,7 +124,7 @@ def create_model(
         configuration={},
         preload=True,
         core=None,
-        weights_path=None,
+        weights_path="",
         adaptor_parameters={},
         device="AUTO",
         nstreams="1",
@@ -469,3 +469,20 @@ def log_layers_info(self):
                     name, metadata.shape, metadata.precision, metadata.layout
                 )
             )
+
+    def get_model(self):
+        model = self.inference_adapter.get_model()
+        model.set_rt_info(self.__model__, ["model_info", "model_type"])
+        for name in self.parameters():
+            if [] == getattr(self, name):
+                # ov cant serialize empty list. Replace it with ""
+                # TODO: remove when Anastasia Kuporosova fixes that
+                model.set_rt_info("", ["model_info", name])
+            else:
+                model.set_rt_info(getattr(self, name), ["model_info", name])
+        return model
+
+    def save(self, xml_path, bin_path="", version="UNSPECIFIED"):
+        import openvino.runtime as ov
+
+        ov.serialize(self.get_model(), xml_path, bin_path, version)

model_api/python/openvino/model_api/models/types.py

@@ -103,8 +103,10 @@ def __str__(self) -> str:
 
 
 class StringValue(BaseValue):
-    def __init__(self, choices=(), **kwargs):
-        super().__init__(**kwargs)
+    def __init__(
+        self, choices=(), description="No description available", default_value=""
+    ):
+        super().__init__(description, default_value)
         self.choices = choices
         for choice in self.choices:
             if not isinstance(choice, str):
@@ -161,8 +163,10 @@ def validate(self, value):
 
 
 class ListValue(BaseValue):
-    def __init__(self, value_type=None, **kwargs) -> None:
-        super().__init__(**kwargs)
+    def __init__(
+        self, value_type=None, description="No description available", default_value=[]
+    ) -> None:
+        super().__init__(description, default_value)
         self.value_type = value_type
 
     def from_str(self, value):

tests/python/precommit/test_save.py

@@ -0,0 +1,38 @@
+from openvino.model_api.models import Model
+
+
+def test_detector_save(tmp_path):
+    downloaded = Model.create_model(
+        "ssd300", configuration={"mean_values": [0, 0, 0], "confidence_threshold": 0.6}
+    )
+    xml_path = str(tmp_path / "a.xml")
+    downloaded.save(xml_path)
+    deserialized = Model.create_model(xml_path)
+    assert type(downloaded) == type(deserialized)
+    for attr in downloaded.parameters():
+        assert getattr(downloaded, attr) == getattr(deserialized, attr)
+
+
+def test_classifier_save(tmp_path):
+    downloaded = Model.create_model(
+        "efficientnet-b0-pytorch", configuration={"scale_values": [1, 1, 1], "topk": 6}
+    )
+    xml_path = str(tmp_path / "a.xml")
+    downloaded.save(xml_path)
+    deserialized = Model.create_model(xml_path)
+    assert type(downloaded) == type(deserialized)
+    for attr in downloaded.parameters():
+        assert getattr(downloaded, attr) == getattr(deserialized, attr)
+
+
+def test_segmentor_save(tmp_path):
+    downloaded = Model.create_model(
+        "hrnet-v2-c1-segmentation",
+        configuration={"reverse_input_channels": True, "labels": ["first", "second"]},
+    )
+    xml_path = str(tmp_path / "a.xml")
+    downloaded.save(xml_path)
+    deserialized = Model.create_model(xml_path)
+    assert type(downloaded) == type(deserialized)
+    for attr in downloaded.parameters():
+        assert getattr(downloaded, attr) == getattr(deserialized, attr)