feat(ml)!: customizable ML settings (#3891)

* consolidated endpoints, added live configuration

* added ml settings to server

* added settings dashboard

* updated deps, fixed typos

* simplified modelconfig

updated tests

* Added ml setting accordion for admin page

updated tests

* merge `clipText` and `clipVision`

* added face distance setting

clarified setting

* add clip mode in request, dropdown for face models

* polished ml settings

updated descriptions

* update clip field on error

* removed unused import

* add description for image classification threshold

* pin safetensors for arm wheel

updated poetry lock

* moved dto

* set model type only in ml repository

* revert form-data package install

use fetch instead of axios

* added slotted description with link

updated facial recognition description

clarified effect of disabling tasks

* validation before model load

* removed unnecessary getconfig call

* added migration

* updated api

updated api

updated api

---------

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

machine-learning/app/models/base.py

@@ -60,16 +60,21 @@ class InferenceModel(ABC):
         self._load(**model_kwargs)
         self._loaded = True
 
-    def predict(self, inputs: Any) -> Any:
+    def predict(self, inputs: Any, **model_kwargs: Any) -> Any:
         if not self._loaded:
             print(f"Loading {self.model_type.value.replace('_', ' ')} model...")
             self.load()
+        if model_kwargs:
+            self.configure(**model_kwargs)
         return self._predict(inputs)
 
     @abstractmethod
     def _predict(self, inputs: Any) -> Any:
         ...
 
+    def configure(self, **model_kwargs: Any) -> None:
+        pass
+
     @abstractmethod
     def _download(self, **model_kwargs: Any) -> None:
         ...

machine-learning/app/models/clip.py

@@ -1,5 +1,6 @@
 import os
 import zipfile
+from io import BytesIO
 from typing import Any, Literal
 
 import onnxruntime as ort
@@ -8,7 +9,7 @@ from clip_server.model.clip import BICUBIC, _convert_image_to_rgb
 from clip_server.model.clip_onnx import _MODELS, _S3_BUCKET_V2, CLIPOnnxModel, download_model
 from clip_server.model.pretrained_models import _VISUAL_MODEL_IMAGE_SIZE
 from clip_server.model.tokenization import Tokenizer
-from PIL.Image import Image
+from PIL import Image
 from torchvision.transforms import CenterCrop, Compose, Normalize, Resize, ToTensor
 
 from ..schemas import ModelType
@@ -74,9 +75,12 @@ class CLIPEncoder(InferenceModel):
             image_size = _VISUAL_MODEL_IMAGE_SIZE[CLIPOnnxModel.get_model_name(self.model_name)]
             self.transform = _transform_pil_image(image_size)
 
-    def _predict(self, image_or_text: Image | str) -> list[float]:
+    def _predict(self, image_or_text: Image.Image | str) -> list[float]:
+        if isinstance(image_or_text, bytes):
+            image_or_text = Image.open(BytesIO(image_or_text))
+
         match image_or_text:
-            case Image():
+            case Image.Image():
                 if self.mode == "text":
                     raise TypeError("Cannot encode image as text-only model")
                 pixel_values = self.transform(image_or_text)

machine-learning/app/models/facial_recognition.py

@@ -9,7 +9,6 @@ from insightface.model_zoo import ArcFaceONNX, RetinaFace
 from insightface.utils.face_align import norm_crop
 from insightface.utils.storage import BASE_REPO_URL, download_file
 
-from ..config import settings
 from ..schemas import ModelType
 from .base import InferenceModel
 
@@ -20,7 +19,7 @@ class FaceRecognizer(InferenceModel):
     def __init__(
         self,
         model_name: str,
-        min_score: float = settings.min_face_score,
+        min_score: float = 0.7,
         cache_dir: Path | str | None = None,
         **model_kwargs: Any,
     ) -> None:
@@ -69,11 +68,13 @@ class FaceRecognizer(InferenceModel):
         )
         self.rec_model.prepare(ctx_id=0)
 
-    def _predict(self, image: cv2.Mat) -> list[dict[str, Any]]:
+    def _predict(self, image: np.ndarray[int, np.dtype[Any]] | bytes) -> list[dict[str, Any]]:
+        if isinstance(image, bytes):
+            image = cv2.imdecode(np.frombuffer(image, np.uint8), cv2.IMREAD_COLOR)
         bboxes, kpss = self.det_model.detect(image)
         if bboxes.size == 0:
             return []
-        assert isinstance(kpss, np.ndarray)
+        assert isinstance(image, np.ndarray) and isinstance(kpss, np.ndarray)
 
         scores = bboxes[:, 4].tolist()
         bboxes = bboxes[:, :4].round().tolist()
@@ -102,3 +103,6 @@ class FaceRecognizer(InferenceModel):
     @property
     def cached(self) -> bool:
         return self.cache_dir.is_dir() and any(self.cache_dir.glob("*.onnx"))
+
+    def configure(self, **model_kwargs: Any) -> None:
+        self.det_model.det_thresh = model_kwargs.get("min_score", self.det_model.det_thresh)

machine-learning/app/models/image_classification.py

@@ -1,13 +1,13 @@
+from io import BytesIO
 from pathlib import Path
 from typing import Any
 
 from huggingface_hub import snapshot_download
 from optimum.onnxruntime import ORTModelForImageClassification
 from optimum.pipelines import pipeline
-from PIL.Image import Image
+from PIL import Image
 from transformers import AutoImageProcessor
 
-from ..config import settings
 from ..schemas import ModelType
 from .base import InferenceModel
 
@@ -18,7 +18,7 @@ class ImageClassifier(InferenceModel):
     def __init__(
         self,
         model_name: str,
-        min_score: float = settings.min_tag_score,
+        min_score: float = 0.9,
         cache_dir: Path | str | None = None,
         **model_kwargs: Any,
     ) -> None:
@@ -56,8 +56,13 @@ class ImageClassifier(InferenceModel):
                 feature_extractor=processor,
             )
 
-    def _predict(self, image: Image) -> list[str]:
+    def _predict(self, image: Image.Image | bytes) -> list[str]:
+        if isinstance(image, bytes):
+            image = Image.open(BytesIO(image))
         predictions: list[dict[str, Any]] = self.model(image)  # type: ignore
         tags = [tag for pred in predictions for tag in pred["label"].split(", ") if pred["score"] >= self.min_score]
 
         return tags
+
+    def configure(self, **model_kwargs: Any) -> None:
+        self.min_score = model_kwargs.get("min_score", self.min_score)