Openvino models (#14283)

* Enable model conversion cache for openvino

* Use openvino directly for onnx embeddings if available

* Don't fail if zmq is busy

frigate/comms/dispatcher.py

@@ -142,10 +142,11 @@ class Dispatcher:
             )
 
         def handle_update_model_state():
-            model = payload["model"]
-            state = payload["state"]
-            self.model_state[model] = ModelStatusTypesEnum[state]
-            self.publish("model_state", json.dumps(self.model_state))
+            if payload:
+                model = payload["model"]
+                state = payload["state"]
+                self.model_state[model] = ModelStatusTypesEnum[state]
+                self.publish("model_state", json.dumps(self.model_state))
 
         def handle_model_state():
             self.publish("model_state", json.dumps(self.model_state.copy()))

frigate/comms/inter_process.py

@@ -65,8 +65,11 @@ class InterProcessRequestor:
 
     def send_data(self, topic: str, data: any) -> any:
         """Sends data and then waits for reply."""
-        self.socket.send_json((topic, data))
-        return self.socket.recv_json()
+        try:
+            self.socket.send_json((topic, data))
+            return self.socket.recv_json()
+        except zmq.ZMQError:
+            return ""
 
     def stop(self) -> None:
         self.socket.close()

frigate/detectors/plugins/openvino.py

@@ -3,6 +3,7 @@ import os
 
 import numpy as np
 import openvino as ov
+import openvino.properties as props
 from pydantic import Field
 from typing_extensions import Literal
 
@@ -34,6 +35,8 @@ class OvDetector(DetectionApi):
             logger.error(f"OpenVino model file {detector_config.model.path} not found.")
             raise FileNotFoundError
 
+        os.makedirs("/config/model_cache/openvino", exist_ok=True)
+        self.ov_core.set_property({props.cache_dir: "/config/model_cache/openvino"})
         self.interpreter = self.ov_core.compile_model(
             model=detector_config.model.path, device_name=detector_config.device
         )

frigate/embeddings/functions/onnx.py

@@ -5,7 +5,6 @@ from io import BytesIO
 from typing import Callable, Dict, List, Optional, Union
 
 import numpy as np
-import onnxruntime as ort
 import requests
 from PIL import Image
 
@@ -19,7 +18,7 @@ from frigate.comms.inter_process import InterProcessRequestor
 from frigate.const import MODEL_CACHE_DIR, UPDATE_MODEL_STATE
 from frigate.types import ModelStatusTypesEnum
 from frigate.util.downloader import ModelDownloader
-from frigate.util.model import get_ort_providers
+from frigate.util.model import ONNXModelRunner
 
 warnings.filterwarnings(
     "ignore",
@@ -54,16 +53,12 @@ class GenericONNXEmbedding:
         self.download_urls = download_urls
         self.embedding_function = embedding_function
         self.model_type = model_type  # 'text' or 'vision'
-        self.providers, self.provider_options = get_ort_providers(
-            force_cpu=device == "CPU",
-            requires_fp16=model_size == "large" or self.model_type == "text",
-            openvino_device=device,
-        )
-
+        self.model_size = model_size
+        self.device = device
         self.download_path = os.path.join(MODEL_CACHE_DIR, self.model_name)
         self.tokenizer = None
         self.feature_extractor = None
-        self.session = None
+        self.runner = None
         files_names = list(self.download_urls.keys()) + (
             [self.tokenizer_file] if self.tokenizer_file else []
         )
@@ -124,15 +119,17 @@ class GenericONNXEmbedding:
             )
 
     def _load_model_and_tokenizer(self):
-        if self.session is None:
+        if self.runner is None:
             if self.downloader:
                 self.downloader.wait_for_download()
             if self.model_type == "text":
                 self.tokenizer = self._load_tokenizer()
             else:
                 self.feature_extractor = self._load_feature_extractor()
-            self.session = self._load_model(
-                os.path.join(self.download_path, self.model_file)
+            self.runner = ONNXModelRunner(
+                os.path.join(self.download_path, self.model_file),
+                self.device,
+                self.model_size,
             )
 
     def _load_tokenizer(self):
@@ -149,14 +146,6 @@ class GenericONNXEmbedding:
             f"{MODEL_CACHE_DIR}/{self.model_name}",
         )
 
-    def _load_model(self, path: str) -> Optional[ort.InferenceSession]:
-        if os.path.exists(path):
-            return ort.InferenceSession(
-                path, providers=self.providers, provider_options=self.provider_options
-            )
-        else:
-            return None
-
     def _process_image(self, image):
         if isinstance(image, str):
             if image.startswith("http"):
@@ -170,7 +159,7 @@ class GenericONNXEmbedding:
     ) -> List[np.ndarray]:
         self._load_model_and_tokenizer()
 
-        if self.session is None or (
+        if self.runner is None or (
             self.tokenizer is None and self.feature_extractor is None
         ):
             logger.error(
@@ -188,14 +177,14 @@ class GenericONNXEmbedding:
                 images=processed_images, return_tensors="np"
             )
 
-        input_names = [input.name for input in self.session.get_inputs()]
+        input_names = self.runner.get_input_names()
         onnx_inputs = {
             name: processed_inputs[name]
             for name in input_names
             if name in processed_inputs
         }
 
-        outputs = self.session.run(None, onnx_inputs)
+        outputs = self.runner.run(onnx_inputs)
         embeddings = self.embedding_function(outputs)
 
         return [embedding for embedding in embeddings]

frigate/util/model.py

@@ -1,9 +1,16 @@
 """Model Utils"""
 
 import os
+from typing import Any
 
 import onnxruntime as ort
 
+try:
+    import openvino as ov
+except ImportError:
+    # openvino is not included
+    pass
+
 
 def get_ort_providers(
     force_cpu: bool = False, openvino_device: str = "AUTO", requires_fp16: bool = False
@@ -42,3 +49,56 @@ def get_ort_providers(
             options.append({})
 
     return (providers, options)
+
+
+class ONNXModelRunner:
+    """Run onnx models optimally based on available hardware."""
+
+    def __init__(self, model_path: str, device: str, requires_fp16: bool = False):
+        self.model_path = model_path
+        self.ort: ort.InferenceSession = None
+        self.ov: ov.Core = None
+        providers, options = get_ort_providers(device == "CPU", device, requires_fp16)
+
+        if "OpenVINOExecutionProvider" in providers:
+            # use OpenVINO directly
+            self.type = "ov"
+            self.ov = ov.Core()
+            self.ov.set_property(
+                {ov.properties.cache_dir: "/config/model_cache/openvino"}
+            )
+            self.interpreter = self.ov.compile_model(
+                model=model_path, device_name=device
+            )
+        else:
+            # Use ONNXRuntime
+            self.type = "ort"
+            self.ort = ort.InferenceSession(
+                model_path, providers=providers, provider_options=options
+            )
+
+    def get_input_names(self) -> list[str]:
+        if self.type == "ov":
+            input_names = []
+
+            for input in self.interpreter.inputs:
+                input_names.extend(input.names)
+
+            return input_names
+        elif self.type == "ort":
+            return [input.name for input in self.ort.get_inputs()]
+
+    def run(self, input: dict[str, Any]) -> Any:
+        if self.type == "ov":
+            infer_request = self.interpreter.create_infer_request()
+            input_tensor = list(input.values())
+
+            if len(input_tensor) == 1:
+                input_tensor = ov.Tensor(array=input_tensor[0])
+            else:
+                input_tensor = ov.Tensor(array=input_tensor)
+
+            infer_request.infer(input_tensor)
+            return [infer_request.get_output_tensor().data]
+        elif self.type == "ort":
+            return self.ort.run(None, input)