app.py

import gradio as gr
import joblib
import numpy as np
from tensorflow.keras.models import load_model
from huggingface_hub import snapshot_download
import unicodedata

# ====== Mapping nhãn ======
label_map = {
    0: "Tiêu cực",
    1: "Trung lập",
    2: "Tích cực"
}

MAX_LEN = 200

# ====== Load model ======
local_dir = snapshot_download(
    repo_id="phucn001/SentimentAnalysisModels",
    local_dir="./Models"
)
cnn_lstm_model = load_model(f"{local_dir}/CNNLSTM/best_model.h5", compile=False)
cnn_lstm_tokenizer = joblib.load(f"{local_dir}/CNNLSTM/tokenizer.pkl")
cnn_label_encoder = joblib.load(f"{local_dir}/CNNLSTM/label_encoder.pkl")

def normalize_text(text):
    return unicodedata.normalize("NFC", text).strip().lower()

reverse_label_map = {normalize_text(v): k for k, v in label_map.items()}

def predict_cnn_lstm(text):
    seq = cnn_lstm_tokenizer.texts_to_sequences([text])
    padded = np.zeros((1, MAX_LEN))
    padded[0, -len(seq[0]):] = seq[0][:MAX_LEN]

    probs = cnn_lstm_model.predict(padded, verbose=0)[0]

    # reorder theo label_map
    probs_reordered = np.zeros_like(probs)
    for i, label in enumerate(cnn_label_encoder.classes_):
        norm_label = normalize_text(label)
        new_index = reverse_label_map[norm_label]
        probs_reordered[new_index] = probs[i]

    pred = int(np.argmax(probs_reordered))
    return {
        "label": label_map[pred],
        "probabilities": {label_map[i]: float(probs_reordered[i]) for i in range(len(probs_reordered))}
    }

demo = gr.Interface(
    fn=predict_cnn_lstm,
    inputs=gr.Textbox(lines=2, placeholder="Nhập câu bình luận..."),
    outputs="json",
    title="Sentiment Analysis - CNN-LSTM"
)

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