app.py

from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
from transformers import CLIPProcessor, CLIPModel
from PIL import Image
import requests
from io import BytesIO
import torch

# 1. Initialize the App
app = FastAPI()

# 2. Load the Fashion-CLIP Model
# We use 'CLIPModel' directly to get access to the vector layers.
# This runs ONCE when the server starts.
print("Loading Fashion-CLIP Model...")
model_id = "patrickjohncyh/fashion-clip"
model = CLIPModel.from_pretrained(model_id)
processor = CLIPProcessor.from_pretrained(model_id)
print("Model Loaded!")

# 3. Define Input Data Structures
class TextRequest(BaseModel):
    text: str

class ImageRequest(BaseModel):
    image_url: str

# 4. The Home Route (Health Check)
@app.get("/")
def home():
    return {"status": "Online", "model": "Fashion-CLIP"}

# 5. Endpoint: Convert Text to Vector
@app.post("/embed-text")
def embed_text(req: TextRequest):
    try:
        # Process text
        inputs = processor(text=[req.text], return_tensors="pt", padding=True)
        
        # Calculate features
        with torch.no_grad():  # Disable gradient calculation for CPU speed
            text_features = model.get_text_features(**inputs)
        
        # Normalize the vector (Crucial for Cosine Similarity!)
        text_features = text_features / text_features.norm(p=2, dim=-1, keepdim=True)
        
        # Convert to standard Python list
        vector = text_features[0].tolist()
        return {"vector": vector}
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))

# 6. Endpoint: Convert Image URL to Vector
@app.post("/embed-image")
def embed_image(req: ImageRequest):
    try:
        # Download image
        response = requests.get(req.image_url)
        if response.status_code != 200:
            raise HTTPException(status_code=400, detail="Could not download image")
        
        image = Image.open(BytesIO(response.content))

        # Process image
        inputs = processor(images=image, return_tensors="pt", padding=True)
        
        # Calculate features
        with torch.no_grad():  # Disable gradient calculation for CPU speed
            image_features = model.get_image_features(**inputs)
        
        # Normalize the vector (Crucial for Cosine Similarity!)
        image_features = image_features / image_features.norm(p=2, dim=-1, keepdim=True)
        
        # Convert to standard Python list
        vector = image_features[0].tolist()
        return {"vector": vector}
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))