refactor(ml): modularization and styling (#2835)

* basic refactor and styling

* removed batching

* module entrypoint

* removed unused imports

* model superclass,  model cache now in app state

* fixed cache dir and enforced abstract method

---------

Co-authored-by: Alex Tran <alex.tran1502@gmail.com>

machine-learning/Dockerfile

@@ -21,8 +21,8 @@ ENV NODE_ENV=production \
   PYTHONDONTWRITEBYTECODE=1 \
   PYTHONUNBUFFERED=1 \
   PATH="/opt/venv/bin:$PATH" \
-  PYTHONPATH=`pwd`
+  PYTHONPATH=/usr/src
 
 COPY --from=builder /opt/venv /opt/venv
 COPY app .
-ENTRYPOINT ["python", "main.py"]
+ENTRYPOINT ["python", "-m", "app.main"]

machine-learning/app/config.py

@@ -1,5 +1,10 @@
+from pathlib import Path
+
 from pydantic import BaseSettings
 
+from .schemas import ModelType
+
+
 class Settings(BaseSettings):
     cache_folder: str = "/cache"
     classification_model: str = "microsoft/resnet-50"
@@ -15,8 +20,12 @@ class Settings(BaseSettings):
     min_face_score: float = 0.7
 
     class Config(BaseSettings.Config):
-        env_prefix = 'MACHINE_LEARNING_'
+        env_prefix = "MACHINE_LEARNING_"
         case_sensitive = False
 
 
+def get_cache_dir(model_name: str, model_type: ModelType) -> Path:
+    return Path(settings.cache_folder, model_type.value, model_name)
+
+
 settings = Settings()

machine-learning/app/main.py

@@ -1,52 +1,58 @@
 import os
-import io
+from io import BytesIO
 from typing import Any
 
-from cache import ModelCache
-from schemas import (
+import cv2
+import numpy as np
+import uvicorn
+from fastapi import Body, Depends, FastAPI
+from PIL import Image
+
+from .config import settings
+from .models.base import InferenceModel
+from .models.cache import ModelCache
+from .schemas import (
     EmbeddingResponse,
     FaceResponse,
-    TagResponse,
     MessageResponse,
+    ModelType,
+    TagResponse,
     TextModelRequest,
     TextResponse,
 )
-import uvicorn
-from PIL import Image
-from fastapi import FastAPI, HTTPException, Depends, Body
-from models import get_model, run_classification, run_facial_recognition
-from config import settings
-
-_model_cache = None
 
 app = FastAPI()
 
 
 @app.on_event("startup")
 async def startup_event() -> None:
-    global _model_cache
-    _model_cache = ModelCache(ttl=settings.model_ttl, revalidate=True)
+    app.state.model_cache = ModelCache(ttl=settings.model_ttl, revalidate=True)
+    same_clip = settings.clip_image_model == settings.clip_text_model
+    app.state.clip_vision_type = ModelType.CLIP if same_clip else ModelType.CLIP_VISION
+    app.state.clip_text_type = ModelType.CLIP if same_clip else ModelType.CLIP_TEXT
     models = [
-        (settings.classification_model, "image-classification"),
-        (settings.clip_image_model, "clip"),
-        (settings.clip_text_model, "clip"),
-        (settings.facial_recognition_model, "facial-recognition"),
+        (settings.classification_model, ModelType.IMAGE_CLASSIFICATION),
+        (settings.clip_image_model, app.state.clip_vision_type),
+        (settings.clip_text_model, app.state.clip_text_type),
+        (settings.facial_recognition_model, ModelType.FACIAL_RECOGNITION),
     ]
 
     # Get all models
     for model_name, model_type in models:
         if settings.eager_startup:
-            await _model_cache.get_cached_model(model_name, model_type)
+            await app.state.model_cache.get(model_name, model_type)
         else:
-            get_model(model_name, model_type)
+            InferenceModel.from_model_type(model_type, model_name)
+
 
+def dep_pil_image(byte_image: bytes = Body(...)) -> Image.Image:
+    return Image.open(BytesIO(byte_image))
 
-def dep_model_cache():
-    if _model_cache is None:
-        raise HTTPException(status_code=500, detail="Unable to load model.")
 
-def dep_input_image(image: bytes = Body(...)) -> Image:
-    return Image.open(io.BytesIO(image))
+def dep_cv_image(byte_image: bytes = Body(...)) -> cv2.Mat:
+    byte_image_np = np.frombuffer(byte_image, np.uint8)
+    return cv2.imdecode(byte_image_np, cv2.IMREAD_COLOR)
+
 
 @app.get("/", response_model=MessageResponse)
 async def root() -> dict[str, str]:
@@ -62,33 +68,29 @@ def ping() -> str:
     "/image-classifier/tag-image",
     response_model=TagResponse,
     status_code=200,
-    dependencies=[Depends(dep_model_cache)],
 )
 async def image_classification(
-    image: Image = Depends(dep_input_image)
+    image: Image.Image = Depends(dep_pil_image),
 ) -> list[str]:
-    try:
-        model = await _model_cache.get_cached_model(
-            settings.classification_model, "image-classification"
-        )
-        labels = run_classification(model, image, settings.min_tag_score)
-    except Exception as ex:
-        raise HTTPException(status_code=500, detail=str(ex))
-    else:
-        return labels
+    model = await app.state.model_cache.get(
+        settings.classification_model, ModelType.IMAGE_CLASSIFICATION
+    )
+    labels = model.predict(image)
+    return labels
 
 
 @app.post(
     "/sentence-transformer/encode-image",
     response_model=EmbeddingResponse,
     status_code=200,
-    dependencies=[Depends(dep_model_cache)],
 )
 async def clip_encode_image(
-    image: Image = Depends(dep_input_image)
+    image: Image.Image = Depends(dep_pil_image),
 ) -> list[float]:
-    model = await _model_cache.get_cached_model(settings.clip_image_model, "clip")
-    embedding = model.encode(image).tolist()
+    model = await app.state.model_cache.get(
+        settings.clip_image_model, app.state.clip_vision_type
+    )
+    embedding = model.predict(image)
     return embedding
 
 
@@ -96,13 +98,12 @@ async def clip_encode_image(
     "/sentence-transformer/encode-text",
     response_model=EmbeddingResponse,
     status_code=200,
-    dependencies=[Depends(dep_model_cache)],
 )
-async def clip_encode_text(
-    payload: TextModelRequest
-) -> list[float]:
-    model = await _model_cache.get_cached_model(settings.clip_text_model, "clip")
-    embedding = model.encode(payload.text).tolist()
+async def clip_encode_text(payload: TextModelRequest) -> list[float]:
+    model = await app.state.model_cache.get(
+        settings.clip_text_model, app.state.clip_text_type
+    )
+    embedding = model.predict(payload.text)
     return embedding
 
 
@@ -110,22 +111,21 @@ async def clip_encode_text(
     "/facial-recognition/detect-faces",
     response_model=FaceResponse,
     status_code=200,
-    dependencies=[Depends(dep_model_cache)],
 )
 async def facial_recognition(
-    image: bytes = Body(...),
+    image: cv2.Mat = Depends(dep_cv_image),
 ) -> list[dict[str, Any]]:
-    model = await _model_cache.get_cached_model(
-        settings.facial_recognition_model, "facial-recognition"
+    model = await app.state.model_cache.get(
+        settings.facial_recognition_model, ModelType.FACIAL_RECOGNITION
     )
-    faces = run_facial_recognition(model, image)
+    faces = model.predict(image)
     return faces
 
 
 if __name__ == "__main__":
     is_dev = os.getenv("NODE_ENV") == "development"
     uvicorn.run(
-        "main:app",
+        "app.main:app",
         host=settings.host,
         port=settings.port,
         reload=is_dev,

machine-learning/app/models.py

@@ -1,119 +0,0 @@
-import torch
-from insightface.app import FaceAnalysis
-from pathlib import Path
-
-from transformers import pipeline, Pipeline
-from sentence_transformers import SentenceTransformer
-from typing import Any, BinaryIO
-import cv2 as cv
-import numpy as np
-from PIL import Image
-from config import settings
-
-device = "cuda" if torch.cuda.is_available() else "cpu"
-
-
-def get_model(model_name: str, model_type: str, **model_kwargs):
-    """
-    Instantiates the specified model.
-
-    Args:
-        model_name: Name of model in the model hub used for the task.
-        model_type: Model type or task, which determines which model zoo is used.
-            `facial-recognition` uses Insightface, while all other models use the HF Model Hub.
-
-            Options:
-                `image-classification`, `clip`,`facial-recognition`, `tokenizer`, `processor`
-
-    Returns:
-        model: The requested model.
-    """
-
-    cache_dir = _get_cache_dir(model_name, model_type)
-    match model_type:
-        case "facial-recognition":
-            model = _load_facial_recognition(
-                model_name, cache_dir=cache_dir, **model_kwargs
-            )
-        case "clip":
-            model = SentenceTransformer(
-                model_name, cache_folder=cache_dir, **model_kwargs
-            )
-        case _:
-            model = pipeline(
-                model_type,
-                model_name,
-                model_kwargs={"cache_dir": cache_dir, **model_kwargs},
-            )
-
-    return model
-
-
-def run_classification(
-    model: Pipeline, image: Image, min_score: float | None = None
-):
-    predictions: list[dict[str, Any]] = model(image)  # type: ignore
-    result = {
-        tag
-        for pred in predictions
-        for tag in pred["label"].split(", ")
-        if min_score is None or pred["score"] >= min_score
-    }
-
-    return list(result)
-
-
-def run_facial_recognition(
-    model: FaceAnalysis, image: bytes
-) -> list[dict[str, Any]]:
-    file_bytes = np.frombuffer(image, dtype=np.uint8)
-    img = cv.imdecode(file_bytes, cv.IMREAD_COLOR)
-    height, width, _ = img.shape
-    results = []
-    faces = model.get(img)
-
-    for face in faces:
-        x1, y1, x2, y2 = face.bbox
-
-        results.append(
-            {
-                "imageWidth": width,
-                "imageHeight": height,
-                "boundingBox": {
-                    "x1": round(x1),
-                    "y1": round(y1),
-                    "x2": round(x2),
-                    "y2": round(y2),
-                },
-                "score": face.det_score.item(),
-                "embedding": face.normed_embedding.tolist(),
-            }
-        )
-    return results
-
-
-def _load_facial_recognition(
-    model_name: str,
-    min_face_score: float | None = None,
-    cache_dir: Path | str | None = None,
-    **model_kwargs,
-):
-    if cache_dir is None:
-        cache_dir = _get_cache_dir(model_name, "facial-recognition")
-    if isinstance(cache_dir, Path):
-        cache_dir = cache_dir.as_posix()
-    if min_face_score is None:
-        min_face_score = settings.min_face_score
-
-    model = FaceAnalysis(
-        name=model_name,
-        root=cache_dir,
-        allowed_modules=["detection", "recognition"],
-        **model_kwargs,
-    )
-    model.prepare(ctx_id=0, det_thresh=min_face_score, det_size=(640, 640))
-    return model
-
-
-def _get_cache_dir(model_name: str, model_type: str) -> Path:
-    return Path(settings.cache_folder, device, model_type, model_name)

machine-learning/app/models/__init__.py

@@ -0,0 +1,3 @@
+from .clip import CLIPSTTextEncoder, CLIPSTVisionEncoder
+from .facial_recognition import FaceRecognizer
+from .image_classification import ImageClassifier

machine-learning/app/models/base.py

@@ -0,0 +1,52 @@
+from __future__ import annotations
+
+from abc import abstractmethod, ABC
+from pathlib import Path
+from typing import Any
+
+from ..config import get_cache_dir
+from ..schemas import ModelType
+
+
+class InferenceModel(ABC):
+    _model_type: ModelType
+
+    def __init__(
+        self,
+        model_name: str,
+        cache_dir: Path | None = None,
+    ):
+        self.model_name = model_name
+        self._cache_dir = (
+            cache_dir
+            if cache_dir is not None
+            else get_cache_dir(model_name, self.model_type)
+        )
+
+    @abstractmethod
+    def predict(self, inputs: Any) -> Any:
+        ...
+
+    @property
+    def model_type(self) -> ModelType:
+        return self._model_type
+
+    @property
+    def cache_dir(self) -> Path:
+        return self._cache_dir
+
+    @cache_dir.setter
+    def cache_dir(self, cache_dir: Path):
+        self._cache_dir = cache_dir
+
+    @classmethod
+    def from_model_type(
+        cls, model_type: ModelType, model_name, **model_kwargs
+    ) -> InferenceModel:
+        subclasses = {
+            subclass._model_type: subclass for subclass in cls.__subclasses__()
+        }
+        if model_type not in subclasses:
+            raise ValueError(f"Unsupported model type: {model_type}")
+
+        return subclasses[model_type](model_name, **model_kwargs)

machine-learning/app/cache.py → machine-learning/app/models/cache.py

@@ -1,8 +1,11 @@
-from aiocache.plugins import TimingPlugin, BasePlugin
+import asyncio
+
 from aiocache.backends.memory import SimpleMemoryCache
 from aiocache.lock import OptimisticLock
-from typing import Any
-from models import get_model
+from aiocache.plugins import BasePlugin, TimingPlugin
+
+from ..schemas import ModelType
+from .base import InferenceModel
 
 
 class ModelCache:
@@ -10,7 +13,7 @@ class ModelCache:
 
     def __init__(
         self,
-        ttl: int | None = None,
+        ttl: float | None = None,
         revalidate: bool = False,
         timeout: int | None = None,
         profiling: bool = False,
@@ -35,9 +38,9 @@ class ModelCache:
             ttl=ttl, timeout=timeout, plugins=plugins, namespace=None
         )
 
-    async def get_cached_model(
-        self, model_name: str, model_type: str, **model_kwargs
-    ) -> Any:
+    async def get(
+        self, model_name: str, model_type: ModelType, **model_kwargs
+    ) -> InferenceModel:
         """
         Args:
             model_name: Name of model in the model hub used for the task.
@@ -47,11 +50,16 @@ class ModelCache:
             model: The requested model.
         """
 
-        key = self.cache.build_key(model_name, model_type)
+        key = self.cache.build_key(model_name, model_type.value)
         model = await self.cache.get(key)
         if model is None:
             async with OptimisticLock(self.cache, key) as lock:
-                model = get_model(model_name, model_type, **model_kwargs)
+                model = await asyncio.get_running_loop().run_in_executor(
+                    None,
+                    lambda: InferenceModel.from_model_type(
+                        model_type, model_name, **model_kwargs
+                    ),
+                )
                 await lock.cas(model, ttl=self.ttl)
         return model
 

machine-learning/app/models/clip.py

@@ -0,0 +1,37 @@
+from pathlib import Path
+
+from PIL.Image import Image
+from sentence_transformers import SentenceTransformer
+
+from ..schemas import ModelType
+from .base import InferenceModel
+
+
+class CLIPSTEncoder(InferenceModel):
+    _model_type = ModelType.CLIP
+
+    def __init__(
+        self,
+        model_name: str,
+        cache_dir: Path | None = None,
+        **model_kwargs,
+    ):
+        super().__init__(model_name, cache_dir)
+        self.model = SentenceTransformer(
+            self.model_name,
+            cache_folder=self.cache_dir.as_posix(),
+            **model_kwargs,
+        )
+
+    def predict(self, image_or_text: Image | str) -> list[float]:
+        return self.model.encode(image_or_text).tolist()
+
+
+# stubs to allow different behavior between the two in the future
+# and handle loading different image and text clip models
+class CLIPSTVisionEncoder(CLIPSTEncoder):
+    _model_type = ModelType.CLIP_VISION
+
+
+class CLIPSTTextEncoder(CLIPSTEncoder):
+    _model_type = ModelType.CLIP_TEXT

machine-learning/app/models/facial_recognition.py

@@ -0,0 +1,59 @@
+from pathlib import Path
+from typing import Any
+
+import cv2
+from insightface.app import FaceAnalysis
+
+from ..config import settings
+from ..schemas import ModelType
+from .base import InferenceModel
+
+
+class FaceRecognizer(InferenceModel):
+    _model_type = ModelType.FACIAL_RECOGNITION
+
+    def __init__(
+        self,
+        model_name: str,
+        min_score: float = settings.min_face_score,
+        cache_dir: Path | None = None,
+        **model_kwargs,
+    ):
+        super().__init__(model_name, cache_dir)
+        self.min_score = min_score
+        model = FaceAnalysis(
+            name=self.model_name,
+            root=self.cache_dir.as_posix(),
+            allowed_modules=["detection", "recognition"],
+            **model_kwargs,
+        )
+        model.prepare(
+            ctx_id=0,
+            det_thresh=self.min_score,
+            det_size=(640, 640),
+        )
+        self.model = model
+
+    def predict(self, image: cv2.Mat) -> list[dict[str, Any]]:
+        height, width, _ = image.shape
+        results = []
+        faces = self.model.get(image)
+
+        for face in faces:
+            x1, y1, x2, y2 = face.bbox
+
+            results.append(
+                {
+                    "imageWidth": width,
+                    "imageHeight": height,
+                    "boundingBox": {
+                        "x1": round(x1),
+                        "y1": round(y1),
+                        "x2": round(x2),
+                        "y2": round(y2),
+                    },
+                    "score": face.det_score.item(),
+                    "embedding": face.normed_embedding.tolist(),
+                }
+            )
+        return results

machine-learning/app/models/image_classification.py

@@ -0,0 +1,40 @@
+from pathlib import Path
+
+from PIL.Image import Image
+from transformers.pipelines import pipeline
+
+from ..config import settings
+from ..schemas import ModelType
+from .base import InferenceModel
+
+
+class ImageClassifier(InferenceModel):
+    _model_type = ModelType.IMAGE_CLASSIFICATION
+
+    def __init__(
+        self,
+        model_name: str,
+        min_score: float = settings.min_tag_score,
+        cache_dir: Path | None = None,
+        **model_kwargs,
+    ):
+        super().__init__(model_name, cache_dir)
+        self.min_score = min_score
+
+        self.model = pipeline(
+            self.model_type.value,
+            self.model_name,
+            model_kwargs={"cache_dir": self.cache_dir, **model_kwargs},
+        )
+
+    def predict(self, image: Image) -> list[str]:
+        predictions = self.model(image)
+        tags = list(
+            {
+                tag
+                for pred in predictions
+                for tag in pred["label"].split(", ")
+                if pred["score"] >= self.min_score
+            }
+        )
+        return tags

machine-learning/app/schemas.py

@@ -1,3 +1,5 @@
+from enum import Enum
+
 from pydantic import BaseModel
 
 
@@ -54,3 +56,11 @@ class Face(BaseModel):
 
 class FaceResponse(BaseModel):
     __root__: list[Face]
+
+
+class ModelType(Enum):
+    IMAGE_CLASSIFICATION = "image-classification"
+    CLIP = "clip"
+    CLIP_VISION = "clip-vision"
+    CLIP_TEXT = "clip-text"
+    FACIAL_RECOGNITION = "facial-recognition"