From 1ec9a60e4117205cff2556ff5ae7989c09395ab8 Mon Sep 17 00:00:00 2001
From: Mert <101130780+mertalev@users.noreply.github.com>
Date: Wed, 23 Oct 2024 08:50:28 -0400
Subject: [PATCH] feat(ml): configurable batch size for facial recognition
(#13689)
* configurable batch size, default openvino to 1
* update docs
* don't add a new dependency for two lines
* fix typing
---
docs/docs/install/environment-variables.md | 35 ++++++++++---------
machine-learning/app/config.py | 5 +++
.../models/facial_recognition/recognition.py | 29 +++++++++------
machine-learning/app/test_main.py | 32 +++++++++++++++--
4 files changed, 70 insertions(+), 31 deletions(-)
diff --git a/docs/docs/install/environment-variables.md b/docs/docs/install/environment-variables.md
index 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.py
index 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 = False
ann_tuning_level: int = 2
preload: PreloadModelData | None = None
+ max_batch_size: MaxBatchSize | None = None
@property
def device_id(self) -> str:
diff --git a/machine-learning/app/models/facial_recognition/recognition.py b/machine-learning/app/models/facial_recognition/recognition.py
index 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 Any
import numpy as np
import onnx
+import onnxruntime as ort
from insightface.model_zoo import ArcFaceONNX
from insightface.utils.face_align import norm_crop
from numpy.typing import NDArray
from onnx.tools.update_model_dims import update_inputs_outputs_dims
from PIL import Image
-from app.config import log
+from app.config import log, settings
from app.models.base import InferenceModel
from app.models.transforms import decode_cv2
from 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_default
def _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 1
diff --git a/machine-learning/app/test_main.py b/machine-learning/app/test_main.py
index 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 None
update_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 None
update_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 == 1
update_dims.assert_not_called()
onnx.load.assert_not_called()
onnx.save.assert_not_called()