import requests
from PIL import Image
from io import BytesIO
import torch
from transformers import CLIPProcessor, CLIPModel
import gradio as gr

# Initialize the model and processor
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32").to(device)
processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")

def get_embedding(image_or_text):
    if image_or_text.startswith(('http:', 'https:')):
        # Image URL
        response = requests.get(image_or_text)
        image = Image.open(BytesIO(response.content))
        inputs = processor(images=image, return_tensors="pt").to(device)
        with torch.no_grad():
            features = model.get_image_features(**inputs).cpu().numpy()
    else:
        # Text input
        inputs = processor(text=[image_or_text], return_tensors="pt", padding=True).to(device)
        with torch.no_grad():
            features = model.get_text_features(**inputs).cpu().numpy()

    return features.flatten().tolist()

# Define the Gradio interface
interface = gr.Interface(fn=get_embedding,
                          inputs="text",
                          outputs="json",
                          title="CLIP Model Embeddings",
                          description="Enter an Image URL or text to get embeddings from CLIP.")

if __name__ == "__main__":
    interface.launch(share=True)