feat: update clip api for deployment

.env

@@ -0,0 +1,11 @@
+# PROJECT INFORMATION
+HOST=0.0.0.0
+PORT=7860
+CORS_HEADERS=["*"]
+CORS_ORIGINS=["http://localhost"]
+
+MODEL_NAME="ViT-B/32"
+DEVICE="cpu" # ["cuda", "cpu"]
+
+INDEX_FILE_PATH="data/faiss-index/index_clip_L01_to_L20.faiss"
+KEYFRAMES_GROUPS_JSON_PATH="data/config/keyframes_groups_L01_to_L20.json"

.env.example

@@ -0,0 +1,11 @@
+# PROJECT INFORMATION
+HOST=0.0.0.0
+PORT=8000
+CORS_HEADERS=["*"]
+CORS_ORIGINS=["http://localhost"]
+
+MODEL_NAME="ViT-B/32"
+DEVICE="cpu" # ["cuda", "cpu"]
+
+INDEX_FILE_PATH="data/faiss-index/index.faiss"
+KEYFRAMES_GROUPS_JSON_PATH="data/config/keyframes_groups.json"

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.faiss filter=lfs diff=lfs merge=lfs -text
+*.json filter=lfs diff=lfs merge=lfs -text

Dockerfile

@@ -0,0 +1,37 @@
+FROM python:3.8-slim
+
+RUN apt-get update && \
+    apt-get install git gsutil -y && \
+    apt clean && \
+    rm -rf /var/cache/apt/*
+
+WORKDIR /code
+
+COPY requirements.txt /code/requirements.txt
+
+# PYTHONDONTWRITEBYTECODE=1: Disables the creation of .pyc files (compiled bytecode)
+# PYTHONUNBUFFERED=1: Disables buffering of the standard output stream
+# PYTHONIOENCODING: specifies the encoding to be used for the standard input, output, and error streams
+ENV PYTHONDONTWRITEBYTECODE=1 \
+    PYTHONUNBUFFERED=1 \
+    PYTHONIOENCODING=utf-8
+
+RUN pip install -U pip && \
+    pip install --no-cache-dir -r /code/requirements.txt
+
+RUN useradd -m -u 1000 user
+
+USER user
+
+ENV HOME=/home/user \
+	PATH=/home/user/.local/bin:$PATH
+
+WORKDIR $HOME/app
+
+COPY --chown=user . $HOME/app
+
+# Download index
+# RUN mkdir ./data/faiss-index/ && \
+#     gsutil -m cp "gs://thangtd1/faiss-index/index_clip_L01_to_L20.faiss" ./data/faiss-index/
+
+CMD python ./src/main.py

data/config/keyframes_groups_L01_to_L20.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f8d2f52efda68fed4a80512ecfe30a90e65663da396b61b8de3db11433cd65f3
+size 17780893

data/faiss-index/index_clip_L01_to_L20.faiss

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c294320f4b8cb934f57f199500477324dd57a2e6445db375f02937e5a2fcf19
+size 413999149

requirements.txt

@@ -0,0 +1,16 @@
+fastapi==0.103.1
+uvicorn==0.23.2
+pydantic-settings==2.0.3
+
+
+# Models
+torch==1.7.1
+torchvision==0.8.2
+ftfy==6.1.1
+regex
+tqdm==4.66.1
+git+https://github.com/openai/CLIP.git@main
+
+# Vector Database
+faiss-cpu
+

src/config.py

@@ -0,0 +1,28 @@
+from pathlib import Path
+
+from pydantic_settings import BaseSettings
+
+FILE = Path(__file__)
+ROOT = FILE.parent.parent
+
+
+class Settings(BaseSettings):
+    # API SETTINGS
+    HOST: str
+    PORT: int
+    CORS_ORIGINS: list
+    CORS_HEADERS: list
+
+    # MODEL SETTINGS
+    MODEL_NAME: str = "ViT-B/32"
+    DEVICE: str = "cpu"
+
+    # FAISS DATABASE SETTINGS
+    INDEX_FILE_PATH: str
+    KEYFRAMES_GROUPS_JSON_PATH: str
+
+    class Config:
+        env_file = ROOT / ".env"
+
+
+settings = Settings()

src/itr/dtb_cursor.py

@@ -0,0 +1,28 @@
+import json
+from functools import lru_cache
+
+import faiss
+
+
+class DatabaseCursor:
+    def __init__(self, index_file_path: str, keyframes_groups_json_path: str):
+        self._load_index(index_file_path)
+        self._load_keyframes_groups_info(keyframes_groups_json_path)
+
+    @lru_cache(maxsize=1)
+    def _load_index(self, index_file_path):
+        self.index = faiss.read_index(index_file_path)
+
+    @lru_cache(maxsize=1)
+    def _load_keyframes_groups_info(self, keyframes_groups_json_path: str):
+        with open(keyframes_groups_json_path) as file:
+            self.keyframes_group_info = json.loads(file.read())
+
+    def kNN_search(self, query_vector: str, topk: int = 10):
+        results = []
+        distances, ids = self.index.search(query_vector, topk)
+        for i in range(len(ids[0])):
+            frame_detail = self.keyframes_group_info[ids[0][i]]
+            frame_detail["distance"] = str(distances[0][i])
+            results.append(frame_detail)
+        return results

src/itr/router.py

@@ -0,0 +1,49 @@
+from fastapi import APIRouter, File, status
+from fastapi.responses import JSONResponse
+from pydantic import BaseModel
+
+from .dtb_cursor import DatabaseCursor
+from .vlm_model import VisionLanguageModel
+
+
+class Item(BaseModel):
+    query_text: str
+    topk: int
+
+
+router = APIRouter()
+
+
+vectordb_cursor = None
+vlm_model = None
+
+
+def init_vectordb(**kargs):
+    # Singleton pattern
+    global vectordb_cursor
+    if vectordb_cursor is None:
+        vectordb_cursor = DatabaseCursor(**kargs)
+
+
+def init_model(**kargs):
+    # Singleton
+    global vlm_model
+    if vlm_model is None:
+        vlm_model = VisionLanguageModel(**kargs)
+
+
+@router.post("/retrieval/image-text")
+async def retrieve(item: Item) -> JSONResponse:
+    try:
+        query_vector = vlm_model.get_embedding(input=item.query_text)
+        search_results = vectordb_cursor.kNN_search(query_vector, item.topk)
+    except Exception:
+        return JSONResponse(
+            status_code=status.HTTP_500_INTERNAL_SERVER_ERROR,
+            content={"message": "Search error"},
+        )
+
+    return JSONResponse(
+        status_code=status.HTTP_200_OK,
+        content={"message": "success", "details": search_results},
+    )

src/itr/vlm_model.py

@@ -0,0 +1,30 @@
+from functools import lru_cache
+from typing import Union
+
+import clip
+from PIL import Image
+
+
+class VisionLanguageModel:
+    def __init__(self, model_name: str = "ViT-B/32", device: str = "cuda"):
+        self._load_model(model_name, device)
+        self.device = device
+
+    @lru_cache(maxsize=1)
+    def _load_model(self, model_name, device: str = "cpu"):
+        self.model, self.processor = clip.load(model_name, device=device)
+
+    def get_embedding(self, input: Union[str, Image.Image]):
+        if isinstance(input, str):
+            tokens = clip.tokenize(input).to(self.device)
+            vector = self.model.encode_text(tokens)
+            vector /= vector.norm(dim=-1, keepdim=True)
+            vector = vector.cpu().detach().numpy().astype("float32")
+            return vector
+        elif isinstance(input, Image.Image):
+            image_input = self.preprocess(input).unsqueeze(0).to(self.device)
+            vector = self.model.encode_image(image_input)
+            vector /= vector.norm(dim=-1, keepdim=True)
+            return vector
+        else:
+            raise Exception("Invalid input type")

src/main.py

@@ -0,0 +1,62 @@
+import torch
+from config import settings
+from fastapi import FastAPI, Request, status
+from fastapi.exceptions import RequestValidationError
+from fastapi.middleware.cors import CORSMiddleware
+from fastapi.responses import JSONResponse, RedirectResponse
+from itr.router import init_model, init_vectordb
+from itr.router import router as router
+
+app = FastAPI(title="Text-to-image Retrieval API")
+
+
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=settings.CORS_ORIGINS,
+    allow_headers=settings.CORS_HEADERS,
+    allow_credentials=True,
+    allow_methods=["*"],
+)
+
+
+@app.exception_handler(RequestValidationError)
+async def validation_exception_handler(request: Request, exc: RequestValidationError):
+    # Get the original 'detail' list of errors
+    details = exc.errors()
+    error_details = []
+
+    for error in details:
+        error_details.append({"error": f"{error['msg']} {str(error['loc'])}"})
+    return JSONResponse(content={"message": error_details})
+
+
+@app.on_event("startup")
+async def startup_event():
+    init_vectordb(
+        index_file_path=settings.INDEX_FILE_PATH,
+        keyframes_groups_json_path=settings.KEYFRAMES_GROUPS_JSON_PATH,
+    )
+    device = (
+        "cuda" if settings.DEVICE == "cuda" and torch.cuda.is_available() else "cpu"
+    )
+    init_model(model_name=settings.MODEL_NAME, device=device)
+
+
+@app.get("/", include_in_schema=False)
+async def root() -> None:
+    return RedirectResponse("/docs")
+
+
+@app.get("/health", status_code=status.HTTP_200_OK, tags=["health"])
+async def perform_healthcheck() -> None:
+    return JSONResponse(content={"message": "success"})
+
+
+app.include_router(router)
+
+
+# Start API
+# if __name__ == "__main__":
+#     import uvicorn
+
+#     uvicorn.run("main:app", host=settings.HOST, port=settings.PORT, reload=True)