app.py

import os
import json
import numpy as np
import tensorflow as tf
import gradio as gr

# --- CONFIGURATION ---
ASSETS_DIR = "model_assets" # Nama folder aset

# --- 1. LOAD MODEL & ASSETS ---
print("--- Loading Model & Assets... ---")
try:
    # Load Model .h5
    model_path = os.path.join(ASSETS_DIR, "model_klasifikasi_aduan.h5")
    model = tf.keras.models.load_model(model_path)
    
    # Load Config (Max Sequence Length)
    with open(os.path.join(ASSETS_DIR, "config.json"), 'r') as f:
        config = json.load(f)
    MAX_SEQUENCE_LENGTH = config.get('max_sequence_length', 100) # Default 100 jika tidak ada
    
    # Load Label Map
    with open(os.path.join(ASSETS_DIR, "label_map.json"), 'r') as f:
        label_map = json.load(f)
    index_to_label = {v: k for k, v in label_map.items()}
    
    # Load Vocabulary
    with open(os.path.join(ASSETS_DIR, "vocabulary.txt"), "r") as f:
        vocabulary = [line.strip() for line in f.readlines()]
    word_to_index = {word: index for index, word in enumerate(vocabulary)}
    
    print("✅ Success: Model Loaded!")
except Exception as e:
    print(f"❌ Error Loading Assets: {e}")
    print("Pastikan folder 'model_assets' berisi file .h5, config, label_map, dan vocabulary.")

# --- 2. PREPROCESSING FUNCTION ---
def preprocess_text(text):
    if not text: return None
    tokens = text.lower().split()
    # Mengubah kata ke angka (1 untuk kata yg tidak dikenal/OOV)
    token_indices = [word_to_index.get(token, 1) for token in tokens]
    
    # Padding sequence
    padded_indices = tf.keras.preprocessing.sequence.pad_sequences(
        [token_indices], 
        maxlen=MAX_SEQUENCE_LENGTH, 
        padding='post', 
        truncating='post'
    )
    return padded_indices

# --- 3. PREDICTION FUNCTION ---
def klasifikasi_aduan(text):
    if not text or len(text.strip()) < 3:
        return "⚠️ Mohon masukkan kalimat aduan yang jelas."
    
    processed_input = preprocess_text(text)
    prediction = model.predict(processed_input)[0]
    
    # Format output untuk Gradio (Dictionary {Label: Confidence})
    results = {}
    for i, score in enumerate(prediction):
        label_name = index_to_label.get(i, f"Label {i}")
        results[label_name] = float(score)
    
    return results

# --- 4. GRADIO INTERFACE ---
# Contoh input agar user langsung paham
examples = [
    ["Jalan raya di kecamatan Wonosobo berlubang parah dan membahayakan pengendara motor."],
    ["Anak saya dipersulit saat mengurus administrasi pindah sekolah, mohon bantuannya."],
    ["Lampu penerangan jalan umum di desa mati total sudah satu bulan belum diperbaiki."],
    ["Ada pungutan liar di lokasi wisata yang tidak sesuai dengan tiket resmi."]
]

demo = gr.Interface(
    fn=klasifikasi_aduan,
    inputs=gr.Textbox(lines=4, placeholder="Ketik aduan Anda di sini..."),
    outputs=gr.Label(num_top_classes=3, label="Hasil Analisis Model"),
    title="🏛️ LaporGub! Complaints Classifier",
    description="A Bi-LSTM deep learning model designed to classify Central Java public complaints into 14 government sectors with 87% F1-score accuracy.",
    examples=examples,
    theme="soft"
)

if __name__ == "__main__":
    demo.launch(share=True)