chore(deps): update machine-learning (#6302)

* chore(deps): update machine-learning

* fix typing, use new lifespan syntax

* wrap in try / finally

* move log

---------

Co-authored-by: renovate[bot] <29139614+renovate[bot]@users.noreply.github.com>
Co-authored-by: mertalev <101130780+mertalev@users.noreply.github.com>

machine-learning/Dockerfile

@@ -1,4 +1,4 @@
-FROM python:3.11-bookworm@sha256:291405e32318285d8913b7b03293777c255fb1e89305c82aa495ac747b0049fe as builder
+FROM python:3.11-bookworm@sha256:497c00ec2cff14316a6859c4e30fc88e7ab1f11dd254fb43b8f4b201ca657596 as builder
 
 ENV PYTHONDONTWRITEBYTECODE=1 \
   PYTHONUNBUFFERED=1 \
@@ -27,7 +27,7 @@ RUN if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
   fi
 
 
-FROM python:3.11-slim-bookworm@sha256:8f64a67710f3d981cf3008d6f9f1dbe61accd7927f165f4e37ea3f8b883ccc3f
+FROM python:3.11-slim-bookworm@sha256:637774748f62b832dc11e7b286e48cd716727ed04b45a0322776c01bc526afc3
 ARG TARGETPLATFORM
 RUN apt-get update && apt-get install -y --no-install-recommends tini libmimalloc2.0 && rm -rf /var/lib/apt/lists/*
 

machine-learning/ann/ann.py

@@ -2,7 +2,7 @@ from __future__ import annotations
 
 from ctypes import CDLL, Array, c_bool, c_char_p, c_int, c_ulong, c_void_p
 from os.path import exists
-from typing import Any, Generic, Protocol, Type, TypeVar
+from typing import Any, Protocol, TypeVar
 
 import numpy as np
 from numpy.typing import NDArray

machine-learning/app/conftest.py

@@ -5,10 +5,10 @@ from unittest import mock
 import numpy as np
 import pytest
 from fastapi.testclient import TestClient
+from numpy.typing import NDArray
 from PIL import Image
 
 from .main import app
-from .schemas import ndarray_f32
 
 
 @pytest.fixture
@@ -17,7 +17,7 @@ def pil_image() -> Image.Image:
 
 
 @pytest.fixture
-def cv_image(pil_image: Image.Image) -> ndarray_f32:
+def cv_image(pil_image: Image.Image) -> NDArray[np.float32]:
     return np.asarray(pil_image)[:, :, ::-1]  # PIL uses RGB while cv2 uses BGR
 
 

machine-learning/app/main.py

@@ -2,11 +2,11 @@ import asyncio
 import gc
 import os
 import signal
-import sys
 import threading
 import time
 from concurrent.futures import ThreadPoolExecutor
-from typing import Any, Iterator
+from contextlib import asynccontextmanager
+from typing import Any, AsyncGenerator, Iterator
 from zipfile import BadZipFile
 
 import orjson
@@ -26,7 +26,6 @@ from .schemas import (
 )
 
 MultiPartParser.max_file_size = 2**26  # spools to disk if payload is 64 MiB or larger
-app = FastAPI()
 
 model_cache = ModelCache(ttl=settings.model_ttl, revalidate=settings.model_ttl > 0)
 thread_pool: ThreadPoolExecutor | None = None
@@ -35,8 +34,8 @@ active_requests = 0
 last_called: float | None = None
 
 
-@app.on_event("startup")
-def startup() -> None:
+@asynccontextmanager
+async def lifespan(_: FastAPI) -> AsyncGenerator[None, None]:
     global thread_pool
     log.info(
         (
@@ -44,15 +43,16 @@ def startup() -> None:
             f"{f'after {settings.model_ttl}s of inactivity' if settings.model_ttl > 0 else 'disabled'}."
         )
     )
+    
+    try:
+        if settings.request_threads > 0:
             # asyncio is a huge bottleneck for performance, so we use a thread pool to run blocking code
             thread_pool = ThreadPoolExecutor(settings.request_threads) if settings.request_threads > 0 else None
+            log.info(f"Initialized request thread pool with {settings.request_threads} threads.")
         if settings.model_ttl > 0 and settings.model_ttl_poll_s > 0:
             asyncio.ensure_future(idle_shutdown_task())
-    log.info(f"Initialized request thread pool with {settings.request_threads} threads.")
-
-
-@app.on_event("shutdown")
-def shutdown() -> None:
+        yield
+    finally:
         log.handlers.clear()
         for model in model_cache.cache._cache.values():
             del model
@@ -71,6 +71,9 @@ def update_state() -> Iterator[None]:
         active_requests -= 1
 
 
+app = FastAPI(lifespan=lifespan)
+
+
 @app.get("/", response_model=MessageResponse)
 async def root() -> dict[str, str]:
     return {"message": "Immich ML"}

machine-learning/app/models/ann.py

@@ -3,10 +3,10 @@ from __future__ import annotations
 from pathlib import Path
 from typing import Any, NamedTuple
 
-from numpy import ascontiguousarray
+import numpy as np
+from numpy.typing import NDArray
 
 from ann.ann import Ann
-from app.schemas import ndarray_f32, ndarray_i32
 
 from ..config import log, settings
 
@@ -56,10 +56,10 @@ class AnnSession:
     def run(
         self,
         output_names: list[str] | None,
-        input_feed: dict[str, ndarray_f32] | dict[str, ndarray_i32],
+        input_feed: dict[str, NDArray[np.float32]] | dict[str, NDArray[np.int32]],
         run_options: Any = None,
-    ) -> list[ndarray_f32]:
-        inputs: list[ndarray_f32] = [ascontiguousarray(v) for v in input_feed.values()]
+    ) -> list[NDArray[np.float32]]:
+        inputs: list[NDArray[np.float32]] = [np.ascontiguousarray(v) for v in input_feed.values()]
         return self.ann.execute(self.model, inputs)
 
 

machine-learning/app/models/clip.py

@@ -6,12 +6,13 @@ from pathlib import Path
 from typing import Any, Literal
 
 import numpy as np
+from numpy.typing import NDArray
 from PIL import Image
 from tokenizers import Encoding, Tokenizer
 
 from app.config import clean_name, log
 from app.models.transforms import crop, get_pil_resampling, normalize, resize, to_numpy
-from app.schemas import ModelType, ndarray_f32, ndarray_i32
+from app.schemas import ModelType
 
 from .base import InferenceModel
 
@@ -40,7 +41,7 @@ class BaseCLIPEncoder(InferenceModel):
             self.vision_model = self._make_session(self.visual_path)
             log.debug(f"Loaded clip vision model '{self.model_name}'")
 
-    def _predict(self, image_or_text: Image.Image | str) -> ndarray_f32:
+    def _predict(self, image_or_text: Image.Image | str) -> NDArray[np.float32]:
         if isinstance(image_or_text, bytes):
             image_or_text = Image.open(BytesIO(image_or_text))
 
@@ -48,7 +49,7 @@ class BaseCLIPEncoder(InferenceModel):
             case Image.Image():
                 if self.mode == "text":
                     raise TypeError("Cannot encode image as text-only model")
-                outputs: ndarray_f32 = self.vision_model.run(None, self.transform(image_or_text))[0][0]
+                outputs: NDArray[np.float32] = self.vision_model.run(None, self.transform(image_or_text))[0][0]
             case str():
                 if self.mode == "vision":
                     raise TypeError("Cannot encode text as vision-only model")
@@ -59,11 +60,11 @@ class BaseCLIPEncoder(InferenceModel):
         return outputs
 
     @abstractmethod
-    def tokenize(self, text: str) -> dict[str, ndarray_i32]:
+    def tokenize(self, text: str) -> dict[str, NDArray[np.int32]]:
         pass
 
     @abstractmethod
-    def transform(self, image: Image.Image) -> dict[str, ndarray_f32]:
+    def transform(self, image: Image.Image) -> dict[str, NDArray[np.float32]]:
         pass
 
     @property
@@ -161,11 +162,11 @@ class OpenCLIPEncoder(BaseCLIPEncoder):
         self.tokenizer.enable_truncation(max_length=context_length)
         log.debug(f"Loaded tokenizer for CLIP model '{self.model_name}'")
 
-    def tokenize(self, text: str) -> dict[str, ndarray_i32]:
+    def tokenize(self, text: str) -> dict[str, NDArray[np.int32]]:
         tokens: Encoding = self.tokenizer.encode(text)
         return {"text": np.array([tokens.ids], dtype=np.int32)}
 
-    def transform(self, image: Image.Image) -> dict[str, ndarray_f32]:
+    def transform(self, image: Image.Image) -> dict[str, NDArray[np.float32]]:
         image = resize(image, self.size)
         image = crop(image, self.size)
         image_np = to_numpy(image)
@@ -174,7 +175,7 @@ class OpenCLIPEncoder(BaseCLIPEncoder):
 
 
 class MCLIPEncoder(OpenCLIPEncoder):
-    def tokenize(self, text: str) -> dict[str, ndarray_i32]:
+    def tokenize(self, text: str) -> dict[str, NDArray[np.int32]]:
         tokens: Encoding = self.tokenizer.encode(text)
         return {
             "input_ids": np.array([tokens.ids], dtype=np.int32),

machine-learning/app/models/facial_recognition.py

@@ -5,9 +5,10 @@ import cv2
 import numpy as np
 from insightface.model_zoo import ArcFaceONNX, RetinaFace
 from insightface.utils.face_align import norm_crop
+from numpy.typing import NDArray
 
 from app.config import clean_name
-from app.schemas import BoundingBox, Face, ModelType, ndarray_f32
+from app.schemas import Face, ModelType, is_ndarray
 
 from .base import InferenceModel
 
@@ -36,22 +37,25 @@ class FaceRecognizer(InferenceModel):
         )
         self.rec_model.prepare(ctx_id=0)
 
-    def _predict(self, image: ndarray_f32 | bytes) -> list[Face]:
+    def _predict(self, image: NDArray[np.uint8] | bytes) -> list[Face]:
         if isinstance(image, bytes):
-            image = cv2.imdecode(np.frombuffer(image, np.uint8), cv2.IMREAD_COLOR)
-        bboxes, kpss = self.det_model.detect(image)
+            decoded_image = cv2.imdecode(np.frombuffer(image, np.uint8), cv2.IMREAD_COLOR)
+        else:
+            decoded_image = image
+        assert is_ndarray(decoded_image, np.uint8)
+        bboxes, kpss = self.det_model.detect(decoded_image)
         if bboxes.size == 0:
             return []
-        assert isinstance(image, np.ndarray) and isinstance(kpss, np.ndarray)
+        assert is_ndarray(kpss, np.float32)
 
         scores = bboxes[:, 4].tolist()
         bboxes = bboxes[:, :4].round().tolist()
 
         results = []
-        height, width, _ = image.shape
+        height, width, _ = decoded_image.shape
         for (x1, y1, x2, y2), score, kps in zip(bboxes, scores, kpss):
-            cropped_img = norm_crop(image, kps)
-            embedding: ndarray_f32 = self.rec_model.get_feat(cropped_img)[0]
+            cropped_img = norm_crop(decoded_image, kps)
+            embedding: NDArray[np.float32] = self.rec_model.get_feat(cropped_img)[0]
             face: Face = {
                 "imageWidth": width,
                 "imageHeight": height,

machine-learning/app/models/transforms.py

@@ -1,8 +1,7 @@
 import numpy as np
+from numpy.typing import NDArray
 from PIL import Image
 
-from app.schemas import ndarray_f32
-
 _PIL_RESAMPLING_METHODS = {resampling.name.lower(): resampling for resampling in Image.Resampling}
 
 
@@ -23,11 +22,13 @@ def crop(img: Image.Image, size: int) -> Image.Image:
     return img.crop((left, upper, right, lower))
 
 
-def to_numpy(img: Image.Image) -> ndarray_f32:
+def to_numpy(img: Image.Image) -> NDArray[np.float32]:
     return np.asarray(img.convert("RGB")).astype(np.float32) / 255.0
 
 
-def normalize(img: ndarray_f32, mean: float | ndarray_f32, std: float | ndarray_f32) -> ndarray_f32:
+def normalize(
+    img: NDArray[np.float32], mean: float | NDArray[np.float32], std: float | NDArray[np.float32]
+) -> NDArray[np.float32]:
     return (img - mean) / std
 
 

machine-learning/app/schemas.py

@@ -1,13 +1,10 @@
 from enum import StrEnum
-from typing import Any, Protocol, TypeAlias, TypedDict, TypeGuard
+from typing import Any, Protocol, TypedDict, TypeGuard
 
 import numpy as np
+import numpy.typing as npt
 from pydantic import BaseModel
 
-ndarray_f32: TypeAlias = np.ndarray[int, np.dtype[np.float32]]
-ndarray_i64: TypeAlias = np.ndarray[int, np.dtype[np.int64]]
-ndarray_i32: TypeAlias = np.ndarray[int, np.dtype[np.int32]]
-
 
 class TextResponse(BaseModel):
     __root__: str
@@ -35,7 +32,7 @@ class HasProfiling(Protocol):
 
 class Face(TypedDict):
     boundingBox: BoundingBox
-    embedding: ndarray_f32
+    embedding: npt.NDArray[np.float32]
     imageWidth: int
     imageHeight: int
     score: float
@@ -43,3 +40,7 @@ class Face(TypedDict):
 
 def has_profiling(obj: Any) -> TypeGuard[HasProfiling]:
     return hasattr(obj, "profiling") and isinstance(obj.profiling, dict)
+
+
+def is_ndarray(obj: Any, dtype: "type[np._DTypeScalar_co]") -> "TypeGuard[npt.NDArray[np._DTypeScalar_co]]":
+    return isinstance(obj, np.ndarray) and obj.dtype == dtype

machine-learning/export/Dockerfile

@@ -1,4 +1,4 @@
-FROM mambaorg/micromamba:bookworm-slim@sha256:a7033851ca71454ed8cfbc958fdd62a72e6fa90f71ccb3aa9b15e67a98f66a1e as builder
+FROM mambaorg/micromamba:bookworm-slim@sha256:e2586bd4097636856cbf725304ce240c968bea4123bece888d9e02ca2572ea2d as builder
 
 ENV NODE_ENV=production \
   TRANSFORMERS_CACHE=/cache \

machine-learning/pyproject.toml

@@ -60,7 +60,6 @@ warn_untyped_fields = true
 line-length = 120
 target-version = "py311"
 select = ["E", "F", "I"]
-ignore = ["F401"]
 
 [tool.ruff.per-file-ignores]
 "test_main.py" = ["F403"]