import gradio as gr
import tensorflow as tf
from transformers import AutoTokenizer
import numpy as np

# Load tokenizer dari Hugging Face repo
tokenizer = AutoTokenizer.from_pretrained("jeanetrixsiee/javo_analisis_sentiment")

# Load model dari file .h5 secara langsung (bukan pakai TFAutoModel)
model = tf.keras.models.load_model("keras_model")  # folder ini HARUS ada di root Spaces

# Pastikan label2id cocok
id2label = {
    0: 'Negative',
    1: 'Neutral',
    2: 'Positive',
    3: 'Very Negative',
    4: 'Very Positive'
}

def predict_sentiment(text):
    inputs = tokenizer(text, return_tensors="tf", truncation=True, padding=True)
    outputs = model(inputs["input_ids"])

    probs = tf.nn.softmax(outputs, axis=1).numpy()[0]
    predicted_index = int(np.argmax(probs))
    predicted_label = id2label[predicted_index]

    prob_dict = {id2label[i]: round(float(probs[i]) * 100, 2) for i in range(len(probs))}
    sorted_probs = dict(sorted(prob_dict.items(), key=lambda x: x[1], reverse=True))

    return f"Prediksi Sentimen: {predicted_label}", sorted_probs

interface = gr.Interface(
    fn=predict_sentiment,
    inputs=gr.Textbox(label="Masukkan Komentar YouTube"),
    outputs=[
        gr.Textbox(label="Hasil Prediksi"),
        gr.Label(label="Probabilitas Tiap Label (%)")
    ],
    title="🔍 Analisis Sentimen Komentar YouTube",
    description="Model ini memprediksi sentimen dari komentar YouTube dalam 5 kategori:\nVery Negative, Negative, Neutral, Positive, Very Positive."
)

interface.launch()