app.py

import clip
import numpy as np
import torch
import gradio as gr
from PIL import Image
import os

# Load the CLIP model
model, preprocess = clip.load("ViT-B/32")
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device).eval()
print(device)

# Define the Business Listing variable
Business_Listing = "Air Guide"

def find_similarity(image1, image2, text_input):
    if image1 is None or image2 is None:
        return None

    image_features = []

    # Preprocess and encode the two images
    for image in [image1, image2]:
        image = preprocess(image).unsqueeze(0).to(device)
        with torch.no_grad():
            image_feature = model.encode_image(image).float()
            image_features.append(image_feature)

    # Prepare input text
    text_tokens = clip.tokenize([text_input]).to(device)
    text_features = model.encode_text(text_tokens).float()

    # Normalize text features
    text_features /= text_features.norm(dim=-1, keepdim=True)

    similarities = []

    # Calculate cosine similarity for each image
    for image_feature in image_features:
        image_feature /= image_feature.norm(dim=-1, keepdim=True)
        similarity = (text_features @ image_feature.T).cpu().detach().numpy()
        similarities.append(similarity[0, 0])

    # Determine which image has a higher similarity to the text
    best_match_index = 0 if similarities[0] > similarities[1] else 1

    return similarities, best_match_index

# Define a Gradio interface
iface = gr.Interface(
    fn=find_similarity,
    inputs=[
        gr.Image(type="pil", label="Image 1"),
        gr.Image(type="pil", label="Image 2"),
        "text"
    ],
    outputs=["text", "number"],
    live=True,
    interpretation="default",
    title="CLIP Model Image-Text Cosine Similarity",
    description="Upload two images and enter text to find their cosine similarity.",
)

iface.launch()