app.py

import gradio as gr
import os
import re
import time
import torch
import torch.nn as nn
from PIL import Image
import pytesseract
from playwright.sync_api import sync_playwright
import asyncio
from transformers import AutoTokenizer
from torchvision import transforms
from torchvision import models
from torchvision.transforms import functional as F
import pandas as pd
from huggingface_hub import hf_hub_download
import warnings
warnings.filterwarnings("ignore")
from pathlib import Path

# --- Setup ---

# Device setup
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained('indobenchmark/indobert-base-p1')

# Image transformation
class ResizePadToSquare:
    def __init__(self, target_size=300):
        self.target_size = target_size

    def __call__(self, img):
        img = img.convert("RGB")
        img.thumbnail((self.target_size, self.target_size), Image.BILINEAR)
        delta_w = self.target_size - img.size[0]
        delta_h = self.target_size - img.size[1]
        padding = (delta_w // 2, delta_h // 2, delta_w - delta_w // 2, delta_h - delta_h // 2)
        img = F.pad(img, padding, fill=0, padding_mode='constant')
        return img

transform = transforms.Compose([
    ResizePadToSquare(300),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], 
                         std=[0.229, 0.224, 0.225]),
])

# Screenshot folder
SCREENSHOT_DIR = "screenshots"
os.makedirs(SCREENSHOT_DIR, exist_ok=True)

# Set Tesseract language
pytesseract.pytesseract.tesseract_cmd = '/usr/bin/tesseract'  # Path to tesseract in Docker
print("Tesseract OCR initialized.")

# --- Model ---

class LateFusionModel(nn.Module):
    def __init__(self, image_model, text_model):
        super(LateFusionModel, self).__init__()
        self.image_model = image_model
        self.text_model = text_model
        self.image_weight = nn.Parameter(torch.tensor(0.5))
        self.text_weight = nn.Parameter(torch.tensor(0.5))

    def forward(self, images, input_ids, attention_mask):
        with torch.no_grad():
            image_logits = self.image_model(images).squeeze(1)
            text_logits = self.text_model(input_ids=input_ids, attention_mask=attention_mask).logits.squeeze(1)

        weights = torch.softmax(torch.stack([self.image_weight, self.text_weight]), dim=0)
        fused_logits = weights[0] * image_logits + weights[1] * text_logits

        return fused_logits, image_logits, text_logits, weights

# Load model
model_path = "models/best_fusion_model.pt"
if os.path.exists(model_path):
    fusion_model = torch.load(model_path, map_location=device, weights_only=False)
else:
    model_path = hf_hub_download(repo_id="azzandr/gambling-fusion-model", filename="best_fusion_model.pt")
    fusion_model = torch.load(model_path, map_location=device, weights_only=False)

fusion_model.to(device)
fusion_model.eval()
print("Fusion model loaded successfully!")

# Load Image-Only Model
# Load image model from state_dict
image_model_path = "models/best_image_model_Adam_lr0.0001_bs32_state_dict.pt"
if os.path.exists(image_model_path):
    image_only_model = models.efficientnet_b3(weights=models.EfficientNet_B3_Weights.DEFAULT)
    num_features = image_only_model.classifier[1].in_features
    image_only_model.classifier = nn.Linear(num_features, 1)
    image_only_model.load_state_dict(torch.load(image_model_path, map_location=device))
    image_only_model.to(device)
    image_only_model.eval()
    print("Image-only model loaded from state_dict successfully!")
else:
    # Download from HuggingFace if local file doesn't exist
    image_model_path = hf_hub_download(repo_id="azzandr/gambling-fusion-model", 
                                      filename="best_image_model_Adam_lr0.0001_bs32_state_dict.pt")
    image_only_model = models.efficientnet_b3(weights=models.EfficientNet_B3_Weights.DEFAULT)
    num_features = image_only_model.classifier[1].in_features
    image_only_model.classifier = nn.Linear(num_features, 1)
    image_only_model.load_state_dict(torch.load(image_model_path, map_location=device))
    image_only_model.to(device)
    image_only_model.eval()
    print("Image-only model loaded from HuggingFace successfully!")

# --- Functions ---
def clean_text(text):
    exceptions = {
        "di", "ke", "ya"
    }
    # ----- BASIC CLEANING -----
    text = re.sub(r"http\S+", "", text)  # Hapus URL
    text = re.sub(r"\n", " ", text)  # Ganti newline dengan spasi
    text = re.sub(r"[^a-zA-Z']", " ", text)  # Hanya sisakan huruf dan apostrof
    text = re.sub(r"\s{2,}", " ", text).strip().lower()  # Hapus spasi ganda, ubah ke lowercase

    # ----- FILTERING -----
    words = text.split()
    filtered_words = [
        w for w in words
        if (len(w) > 2 or w in exceptions)  # Simpan kata >2 huruf atau ada di exceptions
    ]
    text = ' '.join(filtered_words)

    # ----- REMOVE UNWANTED PATTERNS -----
    text = re.sub(r'\b[aeiou]+\b', '', text)  # Hapus kata semua vokal (panjang berapa pun)
    text = re.sub(r'\b[^aeiou\s]+\b', '', text)  # Hapus kata semua konsonan (panjang berapa pun)
    text = re.sub(r'\b\w{20,}\b', '', text)  # Hapus kata sangat panjang (≥20 huruf)
    text = re.sub(r'\s+', ' ', text).strip()  # Bersihkan spasi ekstra

    # check words number
    if len(text.split()) < 5:
        print(f"Cleaned text too short ({len(text.split())} words). Ignoring text.")
        return ""  # empty return to use image-only
    return text

def take_screenshot(url):
    filename = url.replace('https://', '').replace('http://', '').replace('/', '_').replace('.', '_') + '.png'
    filepath = os.path.join(SCREENSHOT_DIR, filename)
    
    try:
        print(f"Taking screenshot with Playwright for URL: {url}")
        with sync_playwright() as p:
            browser = p.chromium.launch()
            page = browser.new_page(viewport={"width": 1280, "height": 800})
            
            # Add timeout and navigation options
            page.set_default_timeout(60000)  # 60 seconds timeout
            
            # Navigate to the URL with wait until options
            page.goto(url, wait_until="networkidle", timeout=60000)
            
            # Wait a bit for dynamic content to load
            page.wait_for_timeout(3000)
            
            # Take full page screenshot
            page.screenshot(path=filepath, full_page=True)
            browser.close()
            
        print(f"Screenshot taken for URL: {url}")
        return filepath
    except Exception as e:
        print(f"Error taking screenshot with Playwright: {e}")
        return None

def resize_if_needed(image_path, max_mb=1, target_width=720):
    file_size = os.path.getsize(image_path) / (1024 * 1024)  # dalam MB
    if file_size > max_mb:
        try:
            with Image.open(image_path) as img:
                width, height = img.size
                if width > target_width:
                    ratio = target_width / float(width)
                    new_height = int((float(height) * float(ratio)))
                    img = img.resize((target_width, new_height), Image.Resampling.LANCZOS)
                    img.save(image_path, optimize=True, quality=85)
                    print(f"Image resized to {target_width}x{new_height}")
        except Exception as e:
            print(f"Resize error: {e}")

def extract_text_from_image(image_path):
    try:
        resize_if_needed(image_path, max_mb=1, target_width=720)
        
        # Use Tesseract OCR with Indonesian language
        text = pytesseract.image_to_string(Image.open(image_path), lang='ind')
        print(f"OCR text extracted with Tesseract: {len(text)} characters")
        
        return text.strip()
    except Exception as e:
        print(f"Tesseract OCR error: {e}")
        return ""

def prepare_data_for_model(image_path, text):
    image = Image.open(image_path)
    image_tensor = transform(image).unsqueeze(0).to(device)

    clean_text_data = clean_text(text)
    encoding = tokenizer.encode_plus(
        clean_text_data,
        add_special_tokens=True,
        max_length=128,
        padding='max_length',
        truncation=True,
        return_tensors='pt'
    )

    input_ids = encoding['input_ids'].to(device)
    attention_mask = encoding['attention_mask'].to(device)

    return image_tensor, input_ids, attention_mask

def predict_single_url(url):
    if not url.startswith(('http://', 'https://')):
        url = 'https://' + url
        
    screenshot_path = take_screenshot(url)
    if not screenshot_path:
        return f"Error: Failed to take screenshot for {url}", None

    text = extract_text_from_image(screenshot_path)

    if not text.strip():  # Jika text kosong
        print(f"No OCR text found for {url}. Using Image-Only Model.")
        image = Image.open(screenshot_path)
        image_tensor = transform(image).unsqueeze(0).to(device)

        with torch.no_grad():
            image_logits = image_only_model(image_tensor).squeeze(1)
            image_probs = torch.sigmoid(image_logits)

            threshold = 0.6
            is_gambling = image_probs[0] > threshold

        label = "Gambling" if is_gambling else "Non-Gambling"
        confidence = image_probs[0].item() if is_gambling else 1 - image_probs[0].item()
        print(f"[Image-Only] URL: {url}")
        print(f"Prediction: {label} | Confidence: {confidence:.2f}\n")
        return label, f"Confidence: {confidence:.2f}"

    else:
        image_tensor, input_ids, attention_mask = prepare_data_for_model(screenshot_path, text)

        with torch.no_grad():
            fused_logits, image_logits, text_logits, weights = fusion_model(image_tensor, input_ids, attention_mask)
            fused_probs = torch.sigmoid(fused_logits)
            image_probs = torch.sigmoid(image_logits)
            text_probs = torch.sigmoid(text_logits)

            threshold = 0.6
            is_gambling = fused_probs[0] > threshold

        label = "Gambling" if is_gambling else "Non-Gambling"
        confidence = fused_probs[0].item() if is_gambling else 1 - fused_probs[0].item()

        # ✨ Log detail
        print(f"[Fusion Model] URL: {url}")
        print(f"Image Model Prediction Probability: {image_probs[0]:.2f}")
        print(f"Text Model Prediction Probability: {text_probs[0]:.2f}")
        print(f"Fusion Final Prediction: {label} | Confidence: {confidence:.2f}\n")

        return label, f"Confidence: {confidence:.2f}"

def predict_batch_urls(file_obj):
    results = []
    content = file_obj.read().decode('utf-8')
    urls = [line.strip() for line in content.splitlines() if line.strip()]
    for url in urls:
        label, confidence = predict_single_url(url)
        results.append({"url": url, "label": label, "confidence": confidence})

    df = pd.DataFrame(results)
    print(f"Batch prediction completed for {len(urls)} URLs.")
    return df

# --- Gradio App ---

with gr.Blocks() as app:
    gr.Markdown("# 🕵️ Gambling Website Detection (URL Based)")
    gr.Markdown("### Using Playwright & Tesseract OCR")

    with gr.Tab("Single URL"):
        url_input = gr.Textbox(label="Enter Website URL")
        predict_button = gr.Button("Predict")
        label_output = gr.Label()
        confidence_output = gr.Textbox(label="Confidence", interactive=False)

        predict_button.click(fn=predict_single_url, inputs=url_input, outputs=[label_output, confidence_output])

    with gr.Tab("Batch URLs"):
        file_input = gr.File(label="Upload .txt file with URLs (one per line)")
        batch_predict_button = gr.Button("Batch Predict")
        batch_output = gr.DataFrame()

        batch_predict_button.click(fn=predict_batch_urls, inputs=file_input, outputs=batch_output)

if __name__ == "__main__":
    app.launch(server_name="0.0.0.0", server_port=7860)