High performance self-hosted photo and video management solution.
TypeScript backup-toolfluttergoogle-photosgoogle-photos-alternativejavascriptmobile-appnestjsnodejsphoto-galleryphotosphotos-managementself-hostedsveltesveltekittypescriptvideos
feat(ml): configurable batch size for facial recognition (#13689)
Mert1ec9a60e4117
diff --git a/docs/docs/install/environment-variables.md b/docs/docs/install/environment-variables.mdindex bb9b4d434c..e86199dc74 100644--- a/docs/docs/install/environment-variables.md+++ b/docs/docs/install/environment-variables.md@@ -148,23 +148,24 @@ Redis (Sentinel) URL example JSON before encoding:## Machine Learning-| Variable | Description | Default | Containers |-| :-------------------------------------------------------- | :-------------------------------------------------------------------------------------------------- | :-----------------------------------: | :--------------- |-| `MACHINE_LEARNING_MODEL_TTL` | Inactivity time (s) before a model is unloaded (disabled if \<= 0) | `300` | machine learning |-| `MACHINE_LEARNING_MODEL_TTL_POLL_S` | Interval (s) between checks for the model TTL (disabled if \<= 0) | `10` | machine learning |-| `MACHINE_LEARNING_CACHE_FOLDER` | Directory where models are downloaded | `/cache` | machine learning |-| `MACHINE_LEARNING_REQUEST_THREADS`<sup>\*1</sup> | Thread count of the request thread pool (disabled if \<= 0) | number of CPU cores | machine learning |-| `MACHINE_LEARNING_MODEL_INTER_OP_THREADS` | Number of parallel model operations | `1` | machine learning |-| `MACHINE_LEARNING_MODEL_INTRA_OP_THREADS` | Number of threads for each model operation | `2` | machine learning |-| `MACHINE_LEARNING_WORKERS`<sup>\*2</sup> | Number of worker processes to spawn | `1` | machine learning |-| `MACHINE_LEARNING_HTTP_KEEPALIVE_TIMEOUT_S`<sup>\*3</sup> | HTTP Keep-alive time in seconds | `2` | machine learning |-| `MACHINE_LEARNING_WORKER_TIMEOUT` | Maximum time (s) of unresponsiveness before a worker is killed | `120` (`300` if using OpenVINO image) | machine learning |-| `MACHINE_LEARNING_PRELOAD__CLIP` | Name of a CLIP model to be preloaded and kept in cache | | machine learning |-| `MACHINE_LEARNING_PRELOAD__FACIAL_RECOGNITION` | Name of a facial recognition model to be preloaded and kept in cache | | machine learning |-| `MACHINE_LEARNING_ANN` | Enable ARM-NN hardware acceleration if supported | `True` | machine learning |-| `MACHINE_LEARNING_ANN_FP16_TURBO` | Execute operations in FP16 precision: increasing speed, reducing precision (applies only to ARM-NN) | `False` | machine learning |-| `MACHINE_LEARNING_ANN_TUNING_LEVEL` | ARM-NN GPU tuning level (1: rapid, 2: normal, 3: exhaustive) | `2` | machine learning |-| `MACHINE_LEARNING_DEVICE_IDS`<sup>\*4</sup> | Device IDs to use in multi-GPU environments | `0` | machine learning |+| Variable | Description | Default | Containers |+| :-------------------------------------------------------- | :-------------------------------------------------------------------------------------------------- | :-----------------------------: | :--------------- |+| `MACHINE_LEARNING_MODEL_TTL` | Inactivity time (s) before a model is unloaded (disabled if \<= 0) | `300` | machine learning |+| `MACHINE_LEARNING_MODEL_TTL_POLL_S` | Interval (s) between checks for the model TTL (disabled if \<= 0) | `10` | machine learning |+| `MACHINE_LEARNING_CACHE_FOLDER` | Directory where models are downloaded | `/cache` | machine learning |+| `MACHINE_LEARNING_REQUEST_THREADS`<sup>\*1</sup> | Thread count of the request thread pool (disabled if \<= 0) | number of CPU cores | machine learning |+| `MACHINE_LEARNING_MODEL_INTER_OP_THREADS` | Number of parallel model operations | `1` | machine learning |+| `MACHINE_LEARNING_MODEL_INTRA_OP_THREADS` | Number of threads for each model operation | `2` | machine learning |+| `MACHINE_LEARNING_WORKERS`<sup>\*2</sup> | Number of worker processes to spawn | `1` | machine learning |+| `MACHINE_LEARNING_HTTP_KEEPALIVE_TIMEOUT_S`<sup>\*3</sup> | HTTP Keep-alive time in seconds | `2` | machine learning |+| `MACHINE_LEARNING_WORKER_TIMEOUT` | Maximum time (s) of unresponsiveness before a worker is killed | `120` (`300` if using OpenVINO) | machine learning |+| `MACHINE_LEARNING_PRELOAD__CLIP` | Name of a CLIP model to be preloaded and kept in cache | | machine learning |+| `MACHINE_LEARNING_PRELOAD__FACIAL_RECOGNITION` | Name of a facial recognition model to be preloaded and kept in cache | | machine learning |+| `MACHINE_LEARNING_ANN` | Enable ARM-NN hardware acceleration if supported | `True` | machine learning |+| `MACHINE_LEARNING_ANN_FP16_TURBO` | Execute operations in FP16 precision: increasing speed, reducing precision (applies only to ARM-NN) | `False` | machine learning |+| `MACHINE_LEARNING_ANN_TUNING_LEVEL` | ARM-NN GPU tuning level (1: rapid, 2: normal, 3: exhaustive) | `2` | machine learning |+| `MACHINE_LEARNING_DEVICE_IDS`<sup>\*4</sup> | Device IDs to use in multi-GPU environments | `0` | machine learning |+| `MACHINE_LEARNING_MAX_BATCH_SIZE__FACIAL_RECOGNITION` | Set the maximum number of faces that will be processed at once by the facial recognition model | None (`1` if using OpenVINO) | machine learning |\*1: It is recommended to begin with this parameter when changing the concurrency levels of the machine learning service and then tune the other ones.diff --git a/machine-learning/app/config.py b/machine-learning/app/config.pyindex 828dee15f0..92799ac692 100644--- a/machine-learning/app/config.py+++ b/machine-learning/app/config.py@@ -19,6 +19,10 @@ class PreloadModelData(BaseModel):facial_recognition: str | None = None+class MaxBatchSize(BaseModel):+ facial_recognition: int | None = None++class Settings(BaseSettings):model_config = SettingsConfigDict(env_prefix="MACHINE_LEARNING_",@@ -41,6 +45,7 @@ class Settings(BaseSettings):ann_fp16_turbo: bool = Falseann_tuning_level: int = 2preload: PreloadModelData | None = None+ max_batch_size: MaxBatchSize | None = None@propertydef device_id(self) -> str:diff --git a/machine-learning/app/models/facial_recognition/recognition.py b/machine-learning/app/models/facial_recognition/recognition.pyindex c060bdd616..dcfb6b530e 100644--- a/machine-learning/app/models/facial_recognition/recognition.py+++ b/machine-learning/app/models/facial_recognition/recognition.py@@ -3,13 +3,14 @@ from typing import Anyimport numpy as npimport onnx+import onnxruntime as ortfrom insightface.model_zoo import ArcFaceONNXfrom insightface.utils.face_align import norm_cropfrom numpy.typing import NDArrayfrom onnx.tools.update_model_dims import update_inputs_outputs_dimsfrom PIL import Image-from app.config import log+from app.config import log, settingsfrom app.models.base import InferenceModelfrom app.models.transforms import decode_cv2from app.schemas import FaceDetectionOutput, FacialRecognitionOutput, ModelFormat, ModelSession, ModelTask, ModelType@@ -22,11 +23,12 @@ class FaceRecognizer(InferenceModel):def __init__(self, model_name: str, min_score: float = 0.7, **model_kwargs: Any) -> None:super().__init__(model_name, **model_kwargs)self.min_score = model_kwargs.pop("minScore", min_score)- self.batch = self.model_format == ModelFormat.ONNX+ max_batch_size = settings.max_batch_size.facial_recognition if settings.max_batch_size else None+ self.batch_size = max_batch_size if max_batch_size else self._batch_size_defaultdef _load(self) -> ModelSession:session = self._make_session(self.model_path)- if self.batch and str(session.get_inputs()[0].shape[0]) != "batch":+ if (not self.batch_size or self.batch_size > 1) and str(session.get_inputs()[0].shape[0]) != "batch":self._add_batch_axis(self.model_path)session = self._make_session(self.model_path)self.model = ArcFaceONNX(@@ -42,18 +44,18 @@ class FaceRecognizer(InferenceModel):return []inputs = decode_cv2(inputs)cropped_faces = self._crop(inputs, faces)- embeddings = self._predict_batch(cropped_faces) if self.batch else self._predict_single(cropped_faces)+ embeddings = self._predict_batch(cropped_faces)return self.postprocess(faces, embeddings)def _predict_batch(self, cropped_faces: list[NDArray[np.uint8]]) -> NDArray[np.float32]:- embeddings: NDArray[np.float32] = self.model.get_feat(cropped_faces)- return embeddings+ if not self.batch_size or len(cropped_faces) <= self.batch_size:+ embeddings: NDArray[np.float32] = self.model.get_feat(cropped_faces)+ return embeddings- def _predict_single(self, cropped_faces: list[NDArray[np.uint8]]) -> NDArray[np.float32]:- embeddings: list[NDArray[np.float32]] = []- for face in cropped_faces:- embeddings.append(self.model.get_feat(face))- return np.concatenate(embeddings, axis=0)+ batch_embeddings: list[NDArray[np.float32]] = []+ for i in range(0, len(cropped_faces), self.batch_size):+ batch_embeddings.append(self.model.get_feat(cropped_faces[i : i + self.batch_size]))+ return np.concatenate(batch_embeddings, axis=0)def postprocess(self, faces: FaceDetectionOutput, embeddings: NDArray[np.float32]) -> FacialRecognitionOutput:return [@@ -77,3 +79,8 @@ class FaceRecognizer(InferenceModel):output_dims = {proto.graph.output[0].name: ["batch"] + static_output_dims}updated_proto = update_inputs_outputs_dims(proto, input_dims, output_dims)onnx.save(updated_proto, model_path)++ @property+ def _batch_size_default(self) -> int | None:+ providers = ort.get_available_providers()+ return None if self.model_format == ModelFormat.ONNX and "OpenVINOExecutionProvider" not in providers else 1diff --git a/machine-learning/app/test_main.py b/machine-learning/app/test_main.pyindex 50ec188aa4..e5cb63997c 100644--- a/machine-learning/app/test_main.py+++ b/machine-learning/app/test_main.py@@ -549,7 +549,7 @@ class TestFaceRecognition:face_recognizer = FaceRecognizer("buffalo_s", cache_dir=path)face_recognizer.load()- assert face_recognizer.batch is True+ assert face_recognizer.batch_size is Noneupdate_dims.assert_called_once_with(proto, {"input.1": ["batch", 3, 224, 224]}, {"output.1": ["batch", 800]})onnx.save.assert_called_once_with(update_dims.return_value, face_recognizer.model_path)@@ -572,7 +572,7 @@ class TestFaceRecognition:face_recognizer = FaceRecognizer("buffalo_s", cache_dir=path)face_recognizer.load()- assert face_recognizer.batch is True+ assert face_recognizer.batch_size is Noneupdate_dims.assert_not_called()onnx.load.assert_not_called()onnx.save.assert_not_called()@@ -596,7 +596,33 @@ class TestFaceRecognition:face_recognizer = FaceRecognizer("buffalo_s", model_format=ModelFormat.ARMNN, cache_dir=path)face_recognizer.load()- assert face_recognizer.batch is False+ assert face_recognizer.batch_size == 1+ update_dims.assert_not_called()+ onnx.load.assert_not_called()+ onnx.save.assert_not_called()++ def test_recognition_does_not_add_batch_axis_for_openvino(+ self, ort_session: mock.Mock, path: mock.Mock, mocker: MockerFixture+ ) -> None:+ onnx = mocker.patch("app.models.facial_recognition.recognition.onnx", autospec=True)+ update_dims = mocker.patch(+ "app.models.facial_recognition.recognition.update_inputs_outputs_dims", autospec=True+ )+ mocker.patch("app.models.base.InferenceModel.download")+ mocker.patch("app.models.facial_recognition.recognition.ArcFaceONNX")+ path.return_value.__truediv__.return_value.__truediv__.return_value.suffix = ".onnx"++ inputs = [SimpleNamespace(name="input.1", shape=("batch", 3, 224, 224))]+ outputs = [SimpleNamespace(name="output.1", shape=("batch", 800))]+ ort_session.return_value.get_inputs.return_value = inputs+ ort_session.return_value.get_outputs.return_value = outputs++ face_recognizer = FaceRecognizer(+ "buffalo_s", model_format=ModelFormat.ARMNN, cache_dir=path, providers=["OpenVINOExecutionProvider"]+ )+ face_recognizer.load()++ assert face_recognizer.batch_size == 1update_dims.assert_not_called()onnx.load.assert_not_called()onnx.save.assert_not_called()