embeddings.py

import os
import pandas as pd
from pinecone import Pinecone, ServerlessSpec
from dotenv import load_dotenv
import requests
from tqdm import tqdm
from PIL import Image
from transformers import AutoProcessor, CLIPModel
import logging
import time

# Logging setup
logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
logger = logging.getLogger(__name__)

# Load environment variables
load_dotenv()

# Pinecone setup
pc = Pinecone(api_key=os.environ.get("PINECONE_API_KEY"))
index_name = "image-index"
if index_name not in pc.list_indexes().names():
    pc.create_index(
        name=index_name,
        metric="cosine",
        dimension=512,
        spec=ServerlessSpec(cloud="aws", region="us-east-1")
    )
    while not pc.describe_index(index_name).status.get("ready", False):
        logger.info("Waiting for index to be ready...")
        time.sleep(1)
unsplash_index = pc.Index(index_name)

# CLIP setup (loaded once)
model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")

# Load dataset
images_df = pd.read_csv("image.csv")[["photo_id", "photo_image_url"]][:500]
total_images = len(images_df)
logger.info(f"Total images to process: {total_images}")

# Sequential processing function
def process_image(row):
    try:
        url = row["photo_image_url"]
        photo_id = row["photo_id"]
        # Download image
        img = Image.open(requests.get(url, stream=True).raw)
        # Generate embeddings
        inputs = processor(images=img, return_tensors="pt")
        image_features = model.get_image_features(**inputs)
        embeddings = image_features.detach().cpu().numpy().flatten().tolist()
        # Upsert to Pinecone
        unsplash_index.upsert(
            vectors=[{
                "id": photo_id,
                "values": embeddings,
                "metadata": {"url": url, "photo_id": photo_id}
            }],
            namespace="image-search-dataset"
        )
        return f"Processed {photo_id}"
    except Exception as e:
        logger.error(f"Error processing {photo_id} with URL {url}: {e}")
        return f"Error {photo_id}"

# Process images sequentially with tqdm
for _, row in tqdm(images_df.iterrows(), total=total_images, desc="Indexing images"):
    result = process_image(row)
    logger.info(result)

logger.info("Indexing complete!")