Mirrors/blakeblackshear.frigate
1
0
Fork 0
mirror of https://github.com/blakeblackshear/frigate.git synced 2025-09-23 17:52:05 +02:00
Code Issues Projects Releases Wiki Activity

Fix ov for LPR (#20117)

Browse Source 
* Check complex model

* Reset state for complex models

* Send arg

* Fix

* Cleanup
This commit is contained in:
Nicolas Mowen 2025-09-17 15:21:57 -06:00 committed by GitHub
parent 26178444f3
commit 1efff67e32
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 26 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
frigate/detectors/detection_runners.py
View File

@ -156,9 +156,10 @@ class CudaGraphRunner(BaseModelRunner):
class OpenVINOModelRunner(BaseModelRunner): class OpenVINOModelRunner(BaseModelRunner):
"""OpenVINO model runner that handles inference efficiently.""" """OpenVINO model runner that handles inference efficiently."""
def __init__(self, model_path: str, device: str, **kwargs): def __init__(self, model_path: str, device: str, complex_model: bool, **kwargs):
self.model_path = model_path self.model_path = model_path
self.device = device self.device = device
self.complex_model = complex_model
if not os.path.isfile(model_path): if not os.path.isfile(model_path):
raise FileNotFoundError(f"OpenVINO model file {model_path} not found.") raise FileNotFoundError(f"OpenVINO model file {model_path} not found.")
@ -180,14 +181,16 @@ class OpenVINOModelRunner(BaseModelRunner):
# Create reusable inference request # Create reusable inference request
self.infer_request = self.compiled_model.create_infer_request() self.infer_request = self.compiled_model.create_infer_request()
self.input_tensor: ov.Tensor | None = None
if not complex_model:
try: try:
input_shape = self.compiled_model.inputs[0].get_shape() input_shape = self.compiled_model.inputs[0].get_shape()
input_element_type = self.compiled_model.inputs[0].get_element_type() input_element_type = self.compiled_model.inputs[0].get_element_type()
self.input_tensor = ov.Tensor(input_element_type, input_shape) self.input_tensor = ov.Tensor(input_element_type, input_shape)
except RuntimeError: except RuntimeError:
# model is complex and has dynamic shape # model is complex and has dynamic shape
self.input_tensor = None pass
def get_input_names(self) -> list[str]: def get_input_names(self) -> list[str]:
"""Get input names for the model.""" """Get input names for the model."""
@ -234,6 +237,15 @@ class OpenVINOModelRunner(BaseModelRunner):
np.copyto(self.input_tensor.data, input_data) np.copyto(self.input_tensor.data, input_data)
self.infer_request.infer(self.input_tensor) self.infer_request.infer(self.input_tensor)
else: else:
if self.complex_model:
try:
# This ensures the model starts with a clean state for each sequence
# Important for RNN models like PaddleOCR recognition
self.infer_request.reset_state()
except Exception:
# this will raise an exception for models with AUTO set as the device
pass
# Multiple inputs case - set each input by name # Multiple inputs case - set each input by name
for input_name, input_data in inputs.items(): for input_name, input_data in inputs.items():
# Find the input by name # Find the input by name
@ -379,7 +391,9 @@ def get_optimized_runner(
return RKNNModelRunner(rknn_path) return RKNNModelRunner(rknn_path)
if device != "CPU" and is_openvino_gpu_npu_available(): if device != "CPU" and is_openvino_gpu_npu_available():
return OpenVINOModelRunner(model_path, device or "AUTO", **kwargs) return OpenVINOModelRunner(
model_path, device or "AUTO", complex_model, **kwargs
)
providers, options = get_ort_providers(device == "CPU", device, **kwargs) providers, options = get_ort_providers(device == "CPU", device, **kwargs)
ortSession = ort.InferenceSession( ortSession = ort.InferenceSession(

4
frigate/detectors/plugins/openvino.py
View File

@ -43,7 +43,9 @@ class OvDetector(DetectionApi):
self.w = detector_config.model.width self.w = detector_config.model.width
self.runner = OpenVINOModelRunner( self.runner = OpenVINOModelRunner(
model_path=detector_config.model.path, device=detector_config.device model_path=detector_config.model.path,
device=detector_config.device,
complex_model=False,
) )
# For dfine models, also pre-allocate target sizes tensor # For dfine models, also pre-allocate target sizes tensor