app.py

import streamlit as st
import pandas as pd
import numpy as np
import pickle
import tensorflow as tf
import neattext.functions as nfx
from tensorflow.keras.models import load_model
from tensorflow.keras.layers import InputLayer, Dense
from PIL import Image

# 1. Set Page Config
st.set_page_config(page_title="Disaster Tweet Detector", page_icon="🚨", layout="centered")

# 2. The Absolute Final Recovery Fix
# We manually define the input shape (5000) to bypass the broken file metadata
@st.cache_resource
def load_assets():
    # This function creates a dummy layer that Keras expects
    def fix_input_layer(**kwargs):
        # We force the shape to be 5000 because that's what your vectorizer produces
        return InputLayer(shape=(5000,))

    # This dummy function ignores the Keras 3 'DTypePolicy'
    def fix_dtype(*args, **kwargs):
        return None

    custom_objects = {
        'InputLayer': fix_input_layer,
        'SafeInputLayer': fix_input_layer,
        'DTypePolicy': fix_dtype,
        # If it complains about Dense again, we use the standard Dense
        'Dense': Dense 
    }
    
    # Load the model ignoring the 'compile' step and using our manual fixes
    model = load_model(
        'disaster_model.h5', 
        custom_objects=custom_objects, 
        compile=False
    )
    
    # Load the vectorizer
    with open('vectorizer.pkl', 'rb') as f:
        vect = pickle.load(f)
    return model, vect

# 3. Preprocessing Function
def clean_text(text):
    text = text.lower()
    text = nfx.remove_hashtags(text)
    text = nfx.remove_urls(text)
    text = nfx.remove_userhandles(text)
    text = nfx.remove_multiple_spaces(text)
    text = nfx.remove_special_characters(text)
    return text

# --- App Interface ---

try:
    model, vect = load_assets()
except Exception as e:
    st.error(f"Critical Deployment Error: {e}")
    st.warning("Manual Override triggered. Please refresh the page.")
    st.stop()

st.title("🚨 Disaster Tweet Detector")
st.markdown("Analyzing social media sentiment for emergency response.")

try:
    img = Image.open('disaster_wordcloud.png')
    st.image(img, caption="Common Disaster Keywords", use_container_width=True)
except:
    pass

st.subheader("Analyze a Tweet")
user_input = st.text_area("Enter tweet text:", placeholder="e.g., Massive earthquake reported!")

if st.button("Predict"):
    if user_input.strip() != "":
        cleaned_msg = clean_text(user_input)
        
        # Convert text to the 5000-feature vector
        vect_text = vect.transform([cleaned_msg]).toarray()
        
        # Prediction logic
        prediction = model.predict(vect_text)
        probability = float(prediction[0][0])
        
        st.divider()
        if probability > 0.5:
            st.error(f"### ⚠️ Real Disaster Detected! ({probability:.2%})")
            st.progress(probability)
        else:
            st.success(f"### ✅ Normal Tweet ({(1-probability):.2%})")
            st.progress(1 - probability)
            
        with st.expander("Technical Details"):
            st.write(f"Processed Text: {cleaned_msg}")
            st.write(f"Vector Shape: {vect_text.shape}")
    else:
        st.warning("Please enter text.")

st.divider()
st.caption("Data Science Portfolio")