app.py

import torch
from transformers import AutoTokenizer, AutoModel
from huggingface_hub import hf_hub_download
from model import IndoBERTMultitask
import gradio as gr

# Load tokenizer dari Hugging Face Hub
tokenizer = AutoTokenizer.from_pretrained("plipustel/idxstockBERT")

# Load base IndoBERT
base_model = AutoModel.from_pretrained("indobenchmark/indobert-base-p1")

# Load model multitask custom
model = IndoBERTMultitask(base_model=base_model, num_event_labels=43)

# Download bobot model dari Hugging Face Hub
model_path = hf_hub_download(repo_id="plipustel/idxstockBERT", filename="pytorch_model.bin")
model.load_state_dict(torch.load(model_path, map_location="cpu"))
model.eval()

# Label mapping
sentiment_map = {0: "Negative", 1: "Netral", 2: "Positive"}

event_map = {
    0: "IPO",
    1: "banking",
    2: "banking sector",
    3: "commodity",
    4: "commodity + stock recommendation",
    5: "company scandal",
    6: "corporate action",
    7: "corporate appointment",
    8: "corporate expansion",
    9: "corporate finance",
    10: "corporate outlook",
    11: "corporate outlook + stock recommendation",
    12: "corporate performance",
    13: "corporate rebranding",
    14: "corporate strategy",
    15: "crypto",
    16: "currency",
    17: "delisting",
    18: "foreign activity",
    19: "global market",
    20: "government policy",
    21: "insider activity",
    22: "interest rate",
    23: "investment product",
    24: "investment strategy",
    25: "ipo",
    26: "logistics expansion",
    27: "macro economy",
    28: "macroeconomy",
    29: "market index",
    30: "market index + foreign activity",
    31: "market index + stock recommendation",
    32: "market outlook",
    33: "market outlook + stock recommendation",
    34: "merger & acquisition",
    35: "politics",
    36: "private placement",
    37: "property",
    38: "regulation",
    39: "sector outlook",
    40: "sector rotation",
    41: "stock recommendation",
    42: "trading activity"
}


idx_labels = [
    "idx_energy", "idx_basic", "idx_indust", "idx_noncyc", "idx_cyclic",
    "idx_health", "idx_finance", "idx_propert", "idx_techno", "idx_infra", "idx_trans"
]

def predict(text):
    with torch.no_grad():
        encoded = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=64)
        out_sent, out_event, out_idx = model(encoded["input_ids"], encoded["attention_mask"])

        # Sentiment
        sent_label = torch.argmax(out_sent, dim=1).item()
        sent_result = sentiment_map[sent_label]

        # Event
        event_label = torch.argmax(out_event, dim=1).item()
        event_result = event_map.get(event_label, "unknown")

        # IDX (multi-label + bullish/bearish status)
        idx_probs = torch.sigmoid(out_idx).squeeze(0)
        idx_result = []

        for i, prob in enumerate(idx_probs):
            status = "Bullish" if prob.item() > 0.5 else "Bearish"
            idx_result.append(f"{idx_labels[i]}: {status} ({prob.item():.2f})")

    return sent_result, event_result, "\n".join(idx_result)


iface = gr.Interface(
    fn=predict,
    inputs=gr.Textbox(lines=2, placeholder="Masukkan headline berita saham..."),
    outputs=[
        gr.Textbox(label="Sentiment"),
        gr.Textbox(label="Event Type"),
        gr.Textbox(label="IDX Sectors"),
    ],
    title="IndoBERT Multitask: Berita Saham",
    description="Masukkan 1 headline berita saham. Model akan memprediksi sentimen, tipe event, dan sektor IDX terkait."
)

iface.launch()