app.py

import streamlit as st
import tensorflow as tf
import numpy as np
from PIL import Image
import requests
from io import BytesIO
from bs4 import BeautifulSoup
import pandas as pd
import os

def download_model(model_url, model_path):
    if not os.path.exists(model_path):
        response = requests.get(model_url)
        with open(model_path, 'wb') as f:
            f.write(response.content)

def load_model(model_path):
    interpreter = tf.lite.Interpreter(model_path=model_path)
    interpreter.allocate_tensors()
    return interpreter

def preprocess_image(image, input_size):
    image = image.convert('RGB')
    image = image.resize((input_size, input_size))
    image_np = np.array(image, dtype=np.float32)
    image_np = np.expand_dims(image_np, axis=0)
    image_np = image_np / 255.0  # Normalize to [0, 1]
    return image_np

def run_inference(interpreter, input_data):
    input_details = interpreter.get_input_details()
    output_details = interpreter.get_output_details()
    
    interpreter.set_tensor(input_details[0]['index'], input_data)
    interpreter.invoke()
    
    output_data_shopping_intent = interpreter.get_tensor(output_details[0]['index'])
    
    return output_data_shopping_intent

def fetch_images_from_url(url):
    response = requests.get(url)
    soup = BeautifulSoup(response.content, 'html.parser')
    img_tags = soup.find_all('img')
    img_urls = [img['src'] for img in img_tags if 'src' in img.attrs]
    return img_urls

def render_intent_bars(labels, percentages):
    for label, percentage in zip(labels, percentages):
        bar_html = f"""
        <div style='display: flex; align-items: center;'>
            <div style='width: 30%; text-align: right; padding-right: 10px;'>{label}</div>
            <div style='width: 70%; display: flex; align-items: center;'>
                <div style='background-color: #007BFF; height: 10px; width: {percentage}%;'></div>
                <div style='padding-left: 10px;'>{percentage:.2f}%</div>
            </div>
        </div>
        """
        st.markdown(bar_html, unsafe_allow_html=True)

def main():
    st.set_page_config(layout="wide")
    st.title("Shopping Intent Classification - SEO by DEJAN")
    
    st.markdown("""
    Multi-label image classification model [extracted from Chrome](https://dejanmarketing.com/product-image-optimisation-with-chromes-convolutional-neural-network/). The model can be deployed in an automated pipeline capable of classifying product images in bulk. Javascript-based website scraping currently unsupported.
    """)
    
    st.write("Enter a URL to fetch and classify all images on the page:")
    
    model_url = "https://huggingface.co/dejanseo/shopping-intent/resolve/main/model.tflite"
    model_path = "model.tflite"
    download_model(model_url, model_path)

    url = st.text_input("Enter URL")

    if url:
        img_urls = fetch_images_from_url(url)
        if img_urls:
            st.write(f"Found {len(img_urls)} images")
            interpreter = load_model(model_path)
            input_details = interpreter.get_input_details()
            input_shape = input_details[0]['shape']
            input_size = input_shape[1]  # assuming square input

            categories = [
                "No Shopping Intent",
                "Apparel",
                "Home Decor",
                "Other"
            ]

            for img_url in img_urls:
                try:
                    response = requests.get(img_url)
                    image = Image.open(BytesIO(response.content))

                    input_data = preprocess_image(image, input_size)
                    output_data_shopping_intent = run_inference(interpreter, input_data)

                    shopping_intent_percentages = (output_data_shopping_intent.flatten() * 100).tolist()
                    
                    col1, col2 = st.columns([1, 3])
                    with col1:
                        st.image(image.resize((224, 224)), width=224)
                    with col2:
                        st.write(f"[URL]({img_url})")
                        render_intent_bars(categories, shopping_intent_percentages)
                    st.write("---")
                except Exception as e:
                    st.write(f"Could not process image {img_url}: {e}")

            st.markdown("""
            Interested in using this in an automated pipeline for bulk image classification?
            Please [book an appointment](https://dejanmarketing.com/conference/) to discuss your needs.
            """)

if __name__ == "__main__":
    main()