app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForTokenClassification
import torch
import os

# Label mapping - urutan HARUS sama dengan training!
# Urutan ini dari file TSV: O muncul pertama, lalu B-Person, I-Person, B-Place, I-Place, B-Organisation, I-Organisation
LABEL_LIST = ['O', 'B-Person', 'I-Person', 'B-Place', 'I-Place', 'B-Organisation', 'I-Organisation']
LABEL_TO_ID = {l: i for i, l in enumerate(LABEL_LIST)}
ID_TO_LABEL = {i: l for l, i in LABEL_TO_ID.items()}

# Base model
BASE_MODEL = "ChristopherA08/IndoELECTRA"
MODEL_PATH = "./model.pth"

print("Loading tokenizer and model...")
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)

# Load model architecture
model = AutoModelForTokenClassification.from_pretrained(
    BASE_MODEL,
    num_labels=len(LABEL_LIST),
    label2id=LABEL_TO_ID,
    id2label=ID_TO_LABEL
)

# Resize embeddings (penting!)
model.resize_token_embeddings(len(tokenizer))

# Load weights dari .pth file
if os.path.exists(MODEL_PATH):
    print(f"Loading weights from {MODEL_PATH}...")
    state_dict = torch.load(MODEL_PATH, map_location="cpu")
    model.load_state_dict(state_dict)
    print(" Model loaded successfully!")
else:
    print(f"  Warning: {MODEL_PATH} not found. Using base model.")

def predict_ner(text):
    """
    Fungsi untuk prediksi Named Entity Recognition
    """
    if not text.strip():
        return " Masukkan teks terlebih dahulu!"
    
    # Tokenisasi
    tokens = text.split()
    inputs = tokenizer(tokens, is_split_into_words=True, return_tensors="pt", truncation=True, max_length=128)
    
    # Prediksi
    with torch.no_grad():
        outputs = model(**inputs)
    
    predictions = torch.argmax(outputs.logits, dim=-1).squeeze().tolist()
    if isinstance(predictions, int):
        predictions = [predictions]
    
    # Extract entities
    word_ids = inputs.word_ids()
    results = []
    prev_word = None
    
    for idx, word_idx in enumerate(word_ids):
        if word_idx is None or word_idx == prev_word:
            continue
        label = model.config.id2label[predictions[idx]]
        results.append((tokens[word_idx], label))
        prev_word = word_idx
    
    # Format output vertikal
    output = ""
    for i, (token, label) in enumerate(results, 1):
        output += f"**{i}.** `{token}` → **{label}**\n\n"
    
    return output


# Contoh teks - DISABLED sementara karena vocab size issue
examples = None

# Buat Gradio Interface dengan Blocks untuk kontrol lebih
with gr.Blocks(title="IndoELECTRA NER") as demo:
    gr.Markdown("# IndoELECTRA Named Entity Recognition")
    gr.Markdown("Model NER untuk Bahasa Indonesia | Dataset: SINGGALANG | Base Model: ChristopherA08/IndoELECTRA")
    
    with gr.Row():
        input_text = gr.Textbox(
            lines=3,
            placeholder="Contoh: Doni sedang menempuh pendidikan di ITERA",
            label="Input Teks"
        )
    
    with gr.Row():
        submit_btn = gr.Button("Submit", variant="primary")
        clear_btn = gr.Button("Clear")
    
    with gr.Row():
        output_text = gr.Markdown(label="Hasil NER")
    
    gr.Markdown("---\n© 2025 NLP Project")
    
    submit_btn.click(fn=predict_ner, inputs=input_text, outputs=output_text)
    clear_btn.click(fn=lambda: ("", ""), inputs=None, outputs=[input_text, output_text])

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