refactor: migrate person repository to kysely (#15242)

* refactor: migrate person repository to kysely

* `asVector` begone

* linting

* fix metadata faces

* update test

---------

Co-authored-by: Alex <alex.tran1502@gmail.com>
Co-authored-by: mertalev <101130780+mertalev@users.noreply.github.com>

machine-learning/app/models/clip/textual.py

@@ -10,7 +10,7 @@ from tokenizers import Encoding, Tokenizer
 
 from app.config import log
 from app.models.base import InferenceModel
-from app.models.transforms import clean_text
+from app.models.transforms import clean_text, serialize_np_array
 from app.schemas import ModelSession, ModelTask, ModelType
 
 
@@ -18,9 +18,9 @@ class BaseCLIPTextualEncoder(InferenceModel):
     depends = []
     identity = (ModelType.TEXTUAL, ModelTask.SEARCH)
 
-    def _predict(self, inputs: str, **kwargs: Any) -> NDArray[np.float32]:
+    def _predict(self, inputs: str, **kwargs: Any) -> str:
         res: NDArray[np.float32] = self.session.run(None, self.tokenize(inputs))[0][0]
-        return res
+        return serialize_np_array(res)
 
     def _load(self) -> ModelSession:
         session = super()._load()

machine-learning/app/models/clip/visual.py

@@ -10,7 +10,15 @@ from PIL import Image
 
 from app.config import log
 from app.models.base import InferenceModel
-from app.models.transforms import crop_pil, decode_pil, get_pil_resampling, normalize, resize_pil, to_numpy
+from app.models.transforms import (
+    crop_pil,
+    decode_pil,
+    get_pil_resampling,
+    normalize,
+    resize_pil,
+    serialize_np_array,
+    to_numpy,
+)
 from app.schemas import ModelSession, ModelTask, ModelType
 
 
@@ -18,10 +26,10 @@ class BaseCLIPVisualEncoder(InferenceModel):
     depends = []
     identity = (ModelType.VISUAL, ModelTask.SEARCH)
 
-    def _predict(self, inputs: Image.Image | bytes, **kwargs: Any) -> NDArray[np.float32]:
+    def _predict(self, inputs: Image.Image | bytes, **kwargs: Any) -> str:
         image = decode_pil(inputs)
         res: NDArray[np.float32] = self.session.run(None, self.transform(image))[0][0]
-        return res
+        return serialize_np_array(res)
 
     @abstractmethod
     def transform(self, image: Image.Image) -> dict[str, NDArray[np.float32]]:

machine-learning/app/models/facial_recognition/recognition.py

@@ -12,7 +12,7 @@ from PIL import Image
 
 from app.config import log, settings
 from app.models.base import InferenceModel
-from app.models.transforms import decode_cv2
+from app.models.transforms import decode_cv2, serialize_np_array
 from app.schemas import FaceDetectionOutput, FacialRecognitionOutput, ModelFormat, ModelSession, ModelTask, ModelType
 
 
@@ -61,7 +61,7 @@ class FaceRecognizer(InferenceModel):
         return [
             {
                 "boundingBox": {"x1": x1, "y1": y1, "x2": x2, "y2": y2},
-                "embedding": embedding,
+                "embedding": serialize_np_array(embedding),
                 "score": score,
             }
             for (x1, y1, x2, y2), embedding, score in zip(faces["boxes"], embeddings, faces["scores"])

machine-learning/app/models/transforms.py

@@ -4,6 +4,7 @@ from typing import IO
 
 import cv2
 import numpy as np
+import orjson
 from numpy.typing import NDArray
 from PIL import Image
 
@@ -69,3 +70,9 @@ def clean_text(text: str, canonicalize: bool = False) -> str:
     if canonicalize:
         text = text.translate(_PUNCTUATION_TRANS).lower()
     return text
+
+
+# this allows the client to use the array as a string without deserializing only to serialize back to a string
+# TODO: use this in a less invasive way
+def serialize_np_array(arr: NDArray[np.float32]) -> str:
+    return orjson.dumps(arr, option=orjson.OPT_SERIALIZE_NUMPY).decode()

machine-learning/app/schemas.py

@@ -79,7 +79,7 @@ class FaceDetectionOutput(TypedDict):
 
 class DetectedFace(TypedDict):
     boundingBox: BoundingBox
-    embedding: npt.NDArray[np.float32]
+    embedding: str
     score: float
 
 

machine-learning/app/test_main.py

@@ -10,6 +10,7 @@ from unittest import mock
 import cv2
 import numpy as np
 import onnxruntime as ort
+import orjson
 import pytest
 from fastapi import HTTPException
 from fastapi.testclient import TestClient
@@ -346,11 +347,11 @@ class TestCLIP:
         mocked.run.return_value = [[self.embedding]]
 
         clip_encoder = OpenClipVisualEncoder("ViT-B-32__openai", cache_dir="test_cache")
-        embedding = clip_encoder.predict(pil_image)
-
-        assert isinstance(embedding, np.ndarray)
-        assert embedding.shape[0] == clip_model_cfg["embed_dim"]
-        assert embedding.dtype == np.float32
+        embedding_str = clip_encoder.predict(pil_image)
+        assert isinstance(embedding_str, str)
+        embedding = orjson.loads(embedding_str)
+        assert isinstance(embedding, list)
+        assert len(embedding) == clip_model_cfg["embed_dim"]
         mocked.run.assert_called_once()
 
     def test_basic_text(
@@ -368,11 +369,11 @@ class TestCLIP:
         mocker.patch("app.models.clip.textual.Tokenizer.from_file", autospec=True)
 
         clip_encoder = OpenClipTextualEncoder("ViT-B-32__openai", cache_dir="test_cache")
-        embedding = clip_encoder.predict("test search query")
-
-        assert isinstance(embedding, np.ndarray)
-        assert embedding.shape[0] == clip_model_cfg["embed_dim"]
-        assert embedding.dtype == np.float32
+        embedding_str = clip_encoder.predict("test search query")
+        assert isinstance(embedding_str, str)
+        embedding = orjson.loads(embedding_str)
+        assert isinstance(embedding, list)
+        assert len(embedding) == clip_model_cfg["embed_dim"]
         mocked.run.assert_called_once()
 
     def test_openclip_tokenizer(
@@ -508,8 +509,11 @@ class TestFaceRecognition:
             assert isinstance(face.get("boundingBox"), dict)
             assert set(face["boundingBox"]) == {"x1", "y1", "x2", "y2"}
             assert all(isinstance(val, np.float32) for val in face["boundingBox"].values())
-            assert isinstance(face.get("embedding"), np.ndarray)
-            assert face["embedding"].shape[0] == 512
+            embedding_str = face.get("embedding")
+            assert isinstance(embedding_str, str)
+            embedding = orjson.loads(embedding_str)
+            assert isinstance(embedding, list)
+            assert len(embedding) == 512
             assert isinstance(face.get("score", None), np.float32)
 
         rec_model.get_feat.assert_called_once()
@@ -880,8 +884,10 @@ class TestPredictionEndpoints:
         actual = response.json()
         assert response.status_code == 200
         assert isinstance(actual, dict)
-        assert isinstance(actual.get("clip", None), list)
-        assert np.allclose(expected, actual["clip"])
+        embedding = actual.get("clip", None)
+        assert isinstance(embedding, str)
+        parsed_embedding = orjson.loads(embedding)
+        assert np.allclose(expected, parsed_embedding)
 
     def test_clip_text_endpoint(self, responses: dict[str, Any], deployed_app: TestClient) -> None:
         expected = responses["clip"]["text"]
@@ -901,8 +907,10 @@ class TestPredictionEndpoints:
         actual = response.json()
         assert response.status_code == 200
         assert isinstance(actual, dict)
-        assert isinstance(actual.get("clip", None), list)
-        assert np.allclose(expected, actual["clip"])
+        embedding = actual.get("clip", None)
+        assert isinstance(embedding, str)
+        parsed_embedding = orjson.loads(embedding)
+        assert np.allclose(expected, parsed_embedding)
 
     def test_face_endpoint(self, pil_image: Image.Image, responses: dict[str, Any], deployed_app: TestClient) -> None:
         byte_image = BytesIO()
@@ -933,5 +941,8 @@ class TestPredictionEndpoints:
 
         for expected_face, actual_face in zip(responses["facial-recognition"], actual["facial-recognition"]):
             assert expected_face["boundingBox"] == actual_face["boundingBox"]
-            assert np.allclose(expected_face["embedding"], actual_face["embedding"])
+            embedding = actual_face.get("embedding", None)
+            assert isinstance(embedding, str)
+            parsed_embedding = orjson.loads(embedding)
+            assert np.allclose(expected_face["embedding"], parsed_embedding)
             assert np.allclose(expected_face["score"], actual_face["score"])