app.py

import torch
import torchvision.transforms as transforms
from PIL import Image
from sklearn.metrics.pairwise import cosine_similarity
import timm
import numpy as np
import gradio as gr

class ImageEmbedder:
    def __init__(self, model_name='vit_base_patch16_224'):
        self.model = timm.create_model(model_name, pretrained=True)
        self.model.head = torch.nn.Identity()  # Remove classification head
        self.model.eval()
        
        self.transform = transforms.Compose([
            transforms.Resize((224, 224)),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
        ])

    def get_embedding(self, image):
        image = image.convert('RGB')
        image_tensor = self.transform(image).unsqueeze(0)
        
        with torch.no_grad():
            embedding = self.model(image_tensor)
        
        return embedding.squeeze().numpy()

def compare_images(image1, image2, similarity_threshold=0.85):
    embedder = ImageEmbedder()

    # Get embeddings
    embedding1 = embedder.get_embedding(image1)
    embedding2 = embedder.get_embedding(image2)

    # Calculate similarity
    similarity = cosine_similarity(embedding1.reshape(1, -1), embedding2.reshape(1, -1))[0][0]

    # Determine if images are similar
    if similarity > similarity_threshold:
        return f"The images are similar. Similarity score: {similarity:.4f}"
    else:
        return f"The images are not similar. Similarity score: {similarity:.4f}"

def main(image1, image2):
    return compare_images(image1, image2)

iface = gr.Interface(
    fn=main, 
    inputs=[gr.Image(type="pil"), gr.Image(type="pil")], 
    outputs="text",
    title="Image Similarity Checker",
    description="Upload two images to check their similarity based on embeddings."
)

iface.launch()