main_RNN_CNN-LSTM.py

# thesis.py
# -*- coding: utf-8 -*-

import pandas as pd
import emoji
import json
import re
import numpy as np
from underthesea import word_tokenize
from tqdm import tqdm
import torch
from torchtext.vocab import Vectors
from sklearn.model_selection import train_test_split
from sklearn.utils import resample
from sklearn.metrics import (
    accuracy_score,
    classification_report,
    precision_score,
    recall_score,
    f1_score,
    confusion_matrix
)
from tensorflow.keras.preprocessing.sequence import pad_sequences
from torch.utils.data import DataLoader, TensorDataset
import torch.nn as nn
import torch.optim as optim
import tensorflow as tf
import os
import joblib

# ========== CÁC HÀM TIỀN XỬ LÝ ==========

def preprocess_sentence(sentence, abbreviations, emoji_mapping):
    """
    Tiền xử lý 1 câu: chuyển thường, thay thế emoji, xóa từ thô tục, 
    ký tự đặc biệt, chuẩn hóa khoảng trắng, v.v.
    """
    sentence = sentence.lower()
    sentence = replace_emojis(sentence, emoji_mapping)
    sentence = remove_profanity(sentence)
    sentence = remove_special_characters(sentence)
    sentence = normalize_whitespace(sentence)
    sentence = replace_abbreviations(sentence, abbreviations)
    sentence = remove_repeated_characters(sentence)
    sentence = replace_numbers(sentence)
    sentence = tokenize_sentence(sentence)
    return sentence

def replace_emojis(sentence, emoji_mapping):
    processed_sentence = []
    for char in sentence:
        if char in emoji_mapping:
            processed_sentence.append(emoji_mapping[char])
        elif not emoji.is_emoji(char):
            processed_sentence.append(char)
    return ''.join(processed_sentence)

def remove_profanity(sentence):
    profane_words = ["loz", "vloz", "vl", "dm", "đm", "clgt", "dmm", "cc", "vc", "đù mé", "vãi"]
    words = sentence.split()
    filtered_words = [word for word in words if word.lower() not in profane_words]
    return ' '.join(filtered_words)

def remove_special_characters(sentence):
    return re.sub(r"[\^\*@#&$%<>~{}|\\]", "", sentence)

def normalize_whitespace(sentence):
    return ' '.join(sentence.split())

def replace_abbreviations(sentence, abbreviations):
    words = sentence.split()
    replaced_words = [
        " ".join(abbreviations[word]) if word in abbreviations else word
        for word in words
    ]
    return ' '.join(replaced_words)

def remove_repeated_characters(sentence):
    # Ví dụ: "đẹp quáaaaaaa" -> "đẹp quá"
    return re.sub(r"(.)\1{2,}", r"\1", sentence)

def replace_numbers(sentence):
    # Thay toàn bộ số bằng token [number]
    return re.sub(r"\d+", "[number]", sentence)

def tokenize_sentence(sentence):
    # Tách từ bằng underthesea
    return ' '.join(word_tokenize(sentence))


# ========== VOCABULARY CLASS ==========

class Vocabulary:
    def __init__(self):
        self.word2id = {}
        self.word2id['<pad>'] = 0
        self.word2id['<unk>'] = 1
        self.unk_id = 1
        self.id2word = {0: '<pad>', 1: '<unk>'}

    def __getitem__(self, word):
        return self.word2id.get(word, self.unk_id)

    def __contains__(self, word):
        return word in self.word2id

    def __len__(self):
        return len(self.word2id)

    def lookup_tokens(self, indices):
        return [self.id2word[idx] for idx in indices]

    def add(self, word):
        if word not in self.word2id:
            idx = len(self.word2id)
            self.word2id[word] = idx
            self.id2word[idx] = word

    @staticmethod
    def tokenize_corpus(corpus):
        tokenized_corpus = []
        for doc in tqdm(corpus, desc="Tokenizing Corpus"):
            tokens = [w.replace(" ", "_") for w in word_tokenize(doc)]
            tokenized_corpus.append(tokens)
        return tokenized_corpus

    def corpus_to_tensor(self, corpus, is_tokenized=False):
        """
        corpus: list các câu (chuỗi) hoặc list các list từ (nếu is_tokenized=True)
        return: list[list[int]], mỗi câu là 1 list gồm các chỉ số token
        """
        tokenized_corpus = (
            self.tokenize_corpus(corpus) if not is_tokenized else corpus
        )
        return [
            [self[token] for token in doc]
            for doc in tokenized_corpus
        ]


# ========== EMOJI MAPPING ==========

emoji_mapping = {
    "😀": "[joy]", "😃": "[joy]", "😄": "[joy]", "😁": "[joy]", "😆": "[joy]", "😅": "[joy]", "😂": "[joy]", "🤣": "[joy]",
    "🙂": "[love]", "🙃": "[love]", "😉": "[love]", "😊": "[love]", "😇": "[love]", "🥰": "[love]", "😍": "[love]",
    "🤩": "[love]", "😘": "[love]", "😗": "[love]", "☺": "[love]", "😚": "[love]", "😙": "[love]",
    "😋": "[satisfaction]", "😛": "[satisfaction]", "😜": "[satisfaction]", "🤪": "[satisfaction]", "😝": "[satisfaction]",
    "🤑": "[satisfaction]",
    "🤐": "[neutral]", "🤨": "[neutral]", "😐": "[neutral]", "😑": "[neutral]", "😶": "[neutral]",
    "😏": "[sarcasm]",
    "😒": "[disappointment]", "🙄": "[disappointment]", "😬": "[disappointment]",
    "😔": "[sadness]", "😪": "[sadness]", "😢": "[sadness]", "😭": "[sadness]", "😥": "[sadness]", "😓": "[sadness]",
    "😩": "[tiredness]", "😫": "[tiredness]", "🥱": "[tiredness]",
    "🤤": "[discomfort]", "🤢": "[discomfort]", "🤮": "[discomfort]", "🤧": "[discomfort]", "🥵": "[discomfort]",
    "🥶": "[discomfort]", "🥴": "[discomfort]", "😵": "[discomfort]", "🤯": "[discomfort]",
    "😕": "[confused]", "😟": "[confused]", "🙁": "[confused]", "☹": "[confused]",
    "😮": "[surprise]", "😯": "[surprise]", "😲": "[surprise]", "😳": "[surprise]", "🥺": "[pleading]",
    "😦": "[fear]", "😧": "[fear]", "😨": "[fear]", "😰": "[fear]", "😱": "[fear]",
    "😖": "[confusion]", "😣": "[confusion]", "😞": "[confusion]",
    "😤": "[anger]", "😡": "[anger]", "😠": "[anger]", "🤬": "[anger]", "😈": "[mischievous]", "👿": "[mischievous]"
}

def load_abbreviations(path):
    with open(path, "r", encoding="utf-8") as f:
        return json.load(f)


# ========== DATA MANAGER ==========

class DataManager:
    def __init__(self, file_path, abbreviations_path, word2vec_path):
        self.file_path = file_path
        self.abbreviations_path = abbreviations_path
        self.word2vec_path = word2vec_path
        self.vocabulary = None
        self.word_embeddings = None
        self.abbreviations = None
        self.load_abbreviations()

    def load_abbreviations(self):
        with open(self.abbreviations_path, "r", encoding="utf-8") as f:
            self.abbreviations = json.load(f)

    def load_word2vec(self):
        """
        Tải vector từ file word2vec, 
        dùng torchtext.Vectors để load embedding pretrained.
        """
        self.word_embeddings = Vectors(
            name=self.word2vec_path, 
            unk_init=torch.Tensor.normal_
        )

    def create_vocab_from_corpus(self, corpus, max_vocab_size=30000):
        """
        Tạo vocabulary từ corpus, chỉ lấy top max_vocab_size từ.
        """
        vocab = Vocabulary()
        from collections import Counter
        counter = Counter()

        for sent in corpus:
            for token in sent.split():
                counter[token] += 1

        most_common = counter.most_common(max_vocab_size)
        for word, _freq in most_common:
            vocab.add(word)

        return vocab

    def preprocess_data(self):
        df = pd.read_excel(self.file_path)
        if "Sentence" not in df.columns:
            raise ValueError("Cột 'Sentence' không tồn tại trong dataset!")

        # Tiền xử lý từng câu
        df["processed_sentence"] = df["Sentence"].apply(
            lambda x: preprocess_sentence(str(x), self.abbreviations, emoji_mapping)
        )

        # Loại những dòng rỗng
        df = df[df["processed_sentence"].str.strip().astype(bool)]

        # Tạo vocab từ chính dữ liệu
        all_sentences = df["processed_sentence"].tolist()
        self.vocabulary = self.create_vocab_from_corpus(all_sentences, max_vocab_size=30000)

        # Load word2vec
        self.load_word2vec()

        return df

    def build_pretrained_embedding_matrix(self, embedding_dim=100):
        """
        Tạo weight_matrix (numpy) (vocab_size x embedding_dim)
        với trọng số pretrained. 
        """
        vocab_size = len(self.vocabulary)
        weight_matrix = np.random.normal(
            scale=0.1, size=(vocab_size, embedding_dim)
        ).astype(np.float32)

        # Copy vector pretrained
        for word, idx in self.vocabulary.word2id.items():
            if word in self.word_embeddings.stoi:
                weight_matrix[idx] = self.word_embeddings.vectors[
                    self.word_embeddings.stoi[word]
                ]

        return weight_matrix

    def split_and_convert(
        self, df, label_column="Emotion", maxlen=400, test_size=0.2,
        for_keras=False, batch_size=32
    ):
        """
        Chia dữ liệu thành train/test. 
        - for_keras=False → return train_loader, test_loader, label_mapping (PyTorch)
        - for_keras=True  → return X_train, X_test, y_train_onehot, y_test_onehot, label_mapping (Keras)
        """
        if label_column not in df.columns:
            raise ValueError(
                f"Cột '{label_column}' không tồn tại. Hiện có: {df.columns.tolist()}"
            )

        # Tạo mapping nhãn -> số
        label_mapping = {label: idx for idx, label in enumerate(df[label_column].unique())}
        df[label_column] = df[label_column].map(label_mapping)
        if df[label_column].isnull().any():
            missing = df[df[label_column].isnull()][label_column].unique()
            raise ValueError(f"Những nhãn cảm xúc sau không có trong label_mapping: {missing}")

        X = df["processed_sentence"].tolist()
        y = df[label_column].tolist()

        # Stratify to maintain class distribution
        X_train, X_test, y_train, y_test = train_test_split(
            X, y, test_size=test_size, random_state=42, stratify=y
        )

        # Convert text -> index
        X_train_ids = self.vocabulary.corpus_to_tensor(X_train, is_tokenized=False)
        X_test_ids  = self.vocabulary.corpus_to_tensor(X_test,  is_tokenized=False)

        # Pad
        X_train_padded = pad_sequences(X_train_ids, maxlen=maxlen, padding='post', truncating='post')
        X_test_padded  = pad_sequences(X_test_ids,  maxlen=maxlen, padding='post', truncating='post')

        print(">>> Debug Split and Convert:")
        print("X_train_padded.shape:", X_train_padded.shape)
        print("X_test_padded.shape: ", X_test_padded.shape)
        print("y_train length:", len(y_train))
        print("y_test length: ", len(y_test))
        print("vocab_size:", len(self.vocabulary))

        if for_keras:
            num_classes = len(label_mapping)
            y_train_onehot = torch.nn.functional.one_hot(
                torch.tensor(y_train), 
                num_classes=num_classes
            ).numpy()
            y_test_onehot  = torch.nn.functional.one_hot(
                torch.tensor(y_test),
                num_classes=num_classes
            ).numpy()

            print("y_train_onehot.shape:", y_train_onehot.shape)
            print("y_test_onehot.shape: ", y_test_onehot.shape)

            return X_train_padded, X_test_padded, y_train_onehot, y_test_onehot, label_mapping
        else:
            # Trả về DataLoader
            X_train_t = torch.tensor(X_train_padded, dtype=torch.long)
            X_test_t  = torch.tensor(X_test_padded,  dtype=torch.long)
            y_train_t = torch.tensor(y_train, dtype=torch.long)
            y_test_t  = torch.tensor(y_test,  dtype=torch.long)

            train_ds = TensorDataset(X_train_t, y_train_t)
            test_ds  = TensorDataset(X_test_t,  y_test_t)

            train_loader = DataLoader(train_ds, batch_size=batch_size, shuffle=True)
            test_loader  = DataLoader(test_ds, batch_size=batch_size, shuffle=False)

            return train_loader, test_loader, label_mapping


# ========== MÔ HÌNH PYTORCH RNN ==========

class SimpleRNN(nn.Module):
    def __init__(self, pretrained_weight, hidden_dim, output_dim, dropout=0.3):
        super(SimpleRNN, self).__init__()
        vocab_size, embedding_dim = pretrained_weight.shape
        # Tạo nn.Embedding từ pretrained_weight
        self.embedding = nn.Embedding.from_pretrained(
            torch.from_numpy(pretrained_weight),
            freeze=False  # True nếu muốn cố định embedding
        )
        self.rnn = nn.LSTM(embedding_dim, hidden_dim, batch_first=True)
        self.dropout = nn.Dropout(dropout)
        self.fc = nn.Linear(hidden_dim, output_dim)

    def forward(self, x):
        embedded = self.dropout(self.embedding(x))
        _, (hidden, _) = self.rnn(embedded)
        hidden = self.dropout(hidden.squeeze(0))
        output = self.fc(hidden)
        return output


def predict_emotion_rnn(model, text, data_manager, label_mapping, device):
    model.eval()
    with torch.no_grad():
        processed_text = preprocess_sentence(text, data_manager.abbreviations, emoji_mapping)
        tokenized_text = data_manager.vocabulary.tokenize_corpus([processed_text])
        text_ids = data_manager.vocabulary.corpus_to_tensor(tokenized_text, is_tokenized=True)
        text_padded = pad_sequences(text_ids, maxlen=400, padding='post', truncating='post')
        text_tensor = torch.tensor(
            text_padded,
            dtype=torch.long
        ).to(device)

        output = model(text_tensor)
        _, predicted = torch.max(output, 1)
        rev_map = {v: k for k, v in label_mapping.items()}
        return rev_map[predicted.item()]


# ========== MÔ HÌNH KERAS CNN-LSTM ==========

def predict_emotion_cnn_lstm(model, text, data_manager, label_mapping):
    processed_text = preprocess_sentence(text, data_manager.abbreviations, emoji_mapping)
    tokenized_text = data_manager.vocabulary.tokenize_corpus([processed_text])
    text_ids = data_manager.vocabulary.corpus_to_tensor(tokenized_text, is_tokenized=True)
    text_padded = pad_sequences(text_ids, maxlen=400, padding='post', truncating='post')
    output = model.predict(text_padded)
    pred = output.argmax(axis=1)[0]
    rev_map = {v: k for k, v in label_mapping.items()}
    return rev_map[pred]


# ========== MAIN ==========

if __name__ == "__main__":
    from keras.models import Model
    from keras.layers import (
        Input, Embedding, Convolution1D, LSTM, Dense, Dropout, Lambda, concatenate
    )
    from keras.optimizers import Adam
    from keras.callbacks import ModelCheckpoint, EarlyStopping

    # -------- ĐƯỜNG DẪN ----------
    file_path = "train.xlsx"
    abbreviations_path = "abbreviations.json"
    word2vec_path = "word2vec_vi_syllables_100dims.txt"
    output_path = "processed.xlsx"

    # Khởi tạo DataManager
    data_manager = DataManager(
        file_path=file_path,
        abbreviations_path=abbreviations_path,
        word2vec_path=word2vec_path
    )

    # 1) Tiền xử lý, tạo vocab, load word2vec
    df = data_manager.preprocess_data()
    print("Trước khi cân bằng lớp (undersampling/oversampling):")
    print(df["Emotion"].value_counts())

    # 2) Cân bằng lớp dữ liệu (Ví dụ: Oversample 'Other' lên 3000)
    # Bạn có thể điều chỉnh theo nhu cầu của mình
    df_enjoyment = df[df["Emotion"] == "Enjoyment"]
    df_other     = df[df["Emotion"] == "Other"]
    df_anger     = df[df["Emotion"] == "Anger"]
    df_sadness   = df[df["Emotion"] == "Sadness"]
    df_disgust   = df[df["Emotion"] == "Disgust"]
    df_fear      = df[df["Emotion"] == "Fear"]
    df_surprise  = df[df["Emotion"] == "Surprise"]

    # Oversample lớp 'Other' lên 3000 (chỉ minh hoạ)
    if len(df_other) < 3000:
        df_other_oversampled = resample(
            df_other, 
            replace=True,
            n_samples=3000,
            random_state=42
        )
    else:
        df_other_oversampled = df_other

    # Giữ nguyên các lớp khác (hoặc oversample tùy ý)
    df_balanced = pd.concat([
        df_enjoyment,
        df_other_oversampled,
        df_anger,
        df_sadness,
        df_disgust,
        df_fear,
        df_surprise
    ], axis=0)

    df_balanced = df_balanced.sample(frac=1, random_state=42).reset_index(drop=True)
    df = df_balanced

    print("\nSau khi cân bằng lớp (demo oversample):")
    print(df["Emotion"].value_counts())

    # Xuất file (nếu muốn)
    df.to_excel(output_path, index=False)

    # ========== TRAIN RNN PYTORCH ==========

    print("\n========== Training PyTorch SimpleRNN ==========")

    # Xây ma trận embedding pretrained
    pretrained_matrix = data_manager.build_pretrained_embedding_matrix(embedding_dim=100)

    # Chia và chuyển đổi dữ liệu thành DataLoader
    train_loader, test_loader, label_mapping = data_manager.split_and_convert(
        df, label_column="Emotion", maxlen=400, test_size=0.2,
        for_keras=False, batch_size=32
    )

    hidden_dim = 128
    output_dim = len(label_mapping)

    model_rnn = SimpleRNN(pretrained_weight=pretrained_matrix,
                          hidden_dim=hidden_dim,
                          output_dim=output_dim,
                          dropout=0.3)
    criterion = nn.CrossEntropyLoss()
    optimizer = optim.Adam(model_rnn.parameters(), lr=1e-3)

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model_rnn.to(device)

    num_epochs = 20
    for epoch in range(num_epochs):
        model_rnn.train()
        epoch_loss = 0
        correct = 0
        total = 0

        for X_batch, y_batch in train_loader:
            X_batch = X_batch.to(device)
            y_batch = y_batch.to(device)

            optimizer.zero_grad()
            preds = model_rnn(X_batch)
            loss = criterion(preds, y_batch)
            loss.backward()
            optimizer.step()

            epoch_loss += loss.item()
            _, pred_label = torch.max(preds, 1)
            correct += (pred_label == y_batch).sum().item()
            total += y_batch.size(0)

        epoch_accuracy = correct / total
        epoch_loss_avg = epoch_loss / len(train_loader)
        print(f"Epoch {epoch+1}/{num_epochs}, "
              f"Loss: {epoch_loss_avg:.4f}, "
              f"Accuracy: {epoch_accuracy:.4f}")

    # Đánh giá trên test set với detailed metrics
    model_rnn.eval()
    test_loss = 0
    correct = 0
    total = 0
    y_true = []
    y_pred = []
    with torch.no_grad():
        for X_batch, y_batch in test_loader:
            X_batch = X_batch.to(device)
            y_batch = y_batch.to(device)
            preds = model_rnn(X_batch)
            loss = criterion(preds, y_batch)
            test_loss += loss.item()

            _, predicted = torch.max(preds, 1)
            correct += (predicted == y_batch).sum().item()
            total += y_batch.size(0)

            y_true.extend(y_batch.cpu().numpy())
            y_pred.extend(predicted.cpu().numpy())

    test_accuracy = accuracy_score(y_true, y_pred)
    test_loss_avg = test_loss / len(test_loader)
    precision_macro = precision_score(y_true, y_pred, average='macro', zero_division=0)
    precision_weighted = precision_score(y_true, y_pred, average='weighted', zero_division=0)
    recall_macro = recall_score(y_true, y_pred, average='macro', zero_division=0)
    recall_weighted = recall_score(y_true, y_pred, average='weighted', zero_division=0)
    f1_macro = f1_score(y_true, y_pred, average='macro', zero_division=0)
    f1_weighted = f1_score(y_true, y_pred, average='weighted', zero_division=0)
    report = classification_report(y_true, y_pred, target_names=label_mapping.keys(), digits=4)
    conf_matrix = confusion_matrix(y_true, y_pred)

    # In các chỉ số
    print(f"\nTest Loss: {test_loss_avg:.4f}, Test Accuracy: {test_accuracy:.4f}")
    print(f"Precision (Macro): {precision_macro:.4f}")
    print(f"Precision (Weighted): {precision_weighted:.4f}")
    print(f"Recall (Macro): {recall_macro:.4f}")
    print(f"Recall (Weighted): {recall_weighted:.4f}")
    print(f"F1-Score (Macro): {f1_macro:.4f}")
    print(f"F1-Score (Weighted): {f1_weighted:.4f}")

    print("\n========== Classification Report ==========")
    print(report)

    print("\n========== Confusion Matrix ==========")
    print(conf_matrix)

    # Lưu báo cáo vào file
    rnn_report_dir = "rnn_emotion_model"
    os.makedirs(rnn_report_dir, exist_ok=True)
    with open(os.path.join(rnn_report_dir, "classification_report.txt"), "w", encoding="utf-8") as f:
        f.write("========== Classification Report ==========\n")
        f.write(report)
        f.write("\n========== Additional Metrics ==========\n")
        f.write(f"Test Loss: {test_loss_avg:.4f}\n")
        f.write(f"Test Accuracy: {test_accuracy:.4f}\n")
        f.write(f"Precision (Macro): {precision_macro:.4f}\n")
        f.write(f"Precision (Weighted): {precision_weighted:.4f}\n")
        f.write(f"Recall (Macro): {recall_macro:.4f}\n")
        f.write(f"Recall (Weighted): {recall_weighted:.4f}\n")
        f.write(f"F1-Score (Macro): {f1_macro:.4f}\n")
        f.write(f"F1-Score (Weighted): {f1_weighted:.4f}\n")
        f.write("\n========== Confusion Matrix ==========\n")
        f.write(np.array2string(conf_matrix))

    print("\n========== Classification Report saved to 'rnn_emotion_model/classification_report.txt' ==========")

    # Lưu mô hình RNN
    torch.save(model_rnn.state_dict(), os.path.join(rnn_report_dir, "simple_rnn.pth"))
    print("========== RNN Model saved to 'rnn_emotion_model/simple_rnn.pth' ==========")

    # ========== TRAIN CNN-LSTM KERAS ==========

    print("\n========== Training CNN-LSTM (Keras) ==========")

    # Tạo embedding pretrained cho Keras
    # Chúng ta có pretrained_matrix (num_vocab x 100)
    # Sẽ truyền vào layer Embedding(..., weights=[...])
    X_train_keras, X_test_keras, y_train_keras, y_test_keras, label_mapping_keras = data_manager.split_and_convert(
        df, label_column="Emotion", maxlen=400, test_size=0.2,
        for_keras=True
    )

    maxlen = 400
    vocab_size, embedding_dim = pretrained_matrix.shape

    # Chuyển pretrained_matrix -> float32 (đảm bảo Keras nhận dạng)
    pretrained_matrix_keras = pretrained_matrix.astype(np.float32)

    input_layer = Input(shape=(maxlen,), dtype='int32', name='main_input')
    emb_layer = Embedding(
        input_dim=vocab_size,
        output_dim=embedding_dim,
        weights=[pretrained_matrix_keras],
        trainable=True  # True hoặc False tùy muốn fine-tune embedding
    )(input_layer)

    def max_1d(X):
        return tf.reduce_max(X, axis=1)

    con3 = Convolution1D(150, kernel_size=3, activation='relu')(emb_layer)
    pool_con3 = Lambda(max_1d, output_shape=(150,))(con3)

    con5 = Convolution1D(150, kernel_size=5, activation='relu')(emb_layer)
    pool_con5 = Lambda(max_1d, output_shape=(150,))(con5)

    lstm_out = LSTM(128, dropout=0.3)(emb_layer)

    merged = concatenate([pool_con3, pool_con5, lstm_out])
    dense  = Dense(100, activation='relu')(merged)
    drop   = Dropout(0.3)(dense)
    output = Dense(output_dim, activation='softmax')(drop)

    model_cnn_lstm = Model(inputs=input_layer, outputs=output)
    model_cnn_lstm.compile(
        loss='categorical_crossentropy', 
        optimizer=Adam(lr=1e-3), 
        metrics=['accuracy']
    )

    checkpoint = ModelCheckpoint(
        'cnn_lstm_best.keras',
        save_best_only=True,
        monitor='val_accuracy',
        mode='max'
    )
    early_stopping = EarlyStopping(
        monitor='val_accuracy',
        patience=5,
        restore_best_weights=True
    )

    history = model_cnn_lstm.fit(
        X_train_keras, y_train_keras,
        validation_data=(X_test_keras, y_test_keras),
        epochs=30,
        batch_size=32,
        callbacks=[checkpoint, early_stopping]
    )

    # Đánh giá trên test set với detailed metrics
    loss, acc = model_cnn_lstm.evaluate(X_test_keras, y_test_keras)
    print(f"CNN-LSTM Test Loss: {loss:.4f}, Test Accuracy: {acc:.4f}")

    # Thu thập dự đoán và tính toán các chỉ số
    y_pred_cnn_lstm = model_cnn_lstm.predict(X_test_keras)
    y_pred_cnn_lstm = np.argmax(y_pred_cnn_lstm, axis=1)
    y_true_cnn_lstm = np.argmax(y_test_keras, axis=1)

    test_accuracy_cnn_lstm = accuracy_score(y_true_cnn_lstm, y_pred_cnn_lstm)
    precision_macro_cnn_lstm = precision_score(y_true_cnn_lstm, y_pred_cnn_lstm, average='macro', zero_division=0)
    precision_weighted_cnn_lstm = precision_score(y_true_cnn_lstm, y_pred_cnn_lstm, average='weighted', zero_division=0)
    recall_macro_cnn_lstm = recall_score(y_true_cnn_lstm, y_pred_cnn_lstm, average='macro', zero_division=0)
    recall_weighted_cnn_lstm = recall_score(y_true_cnn_lstm, y_pred_cnn_lstm, average='weighted', zero_division=0)
    f1_macro_cnn_lstm = f1_score(y_true_cnn_lstm, y_pred_cnn_lstm, average='macro', zero_division=0)
    f1_weighted_cnn_lstm = f1_score(y_true_cnn_lstm, y_pred_cnn_lstm, average='weighted', zero_division=0)
    report_cnn_lstm = classification_report(y_true_cnn_lstm, y_pred_cnn_lstm, target_names=label_mapping.keys(), digits=4)
    conf_matrix_cnn_lstm = confusion_matrix(y_true_cnn_lstm, y_pred_cnn_lstm)

    # In các chỉ số
    print(f"\nCNN-LSTM Test Accuracy: {test_accuracy_cnn_lstm:.4f}")
    print(f"Precision (Macro): {precision_macro_cnn_lstm:.4f}")
    print(f"Precision (Weighted): {precision_weighted_cnn_lstm:.4f}")
    print(f"Recall (Macro): {recall_macro_cnn_lstm:.4f}")
    print(f"Recall (Weighted): {recall_weighted_cnn_lstm:.4f}")
    print(f"F1-Score (Macro): {f1_macro_cnn_lstm:.4f}")
    print(f"F1-Score (Weighted): {f1_weighted_cnn_lstm:.4f}")

    print("\n========== CNN-LSTM Classification Report ==========")
    print(report_cnn_lstm)

    print("\n========== CNN-LSTM Confusion Matrix ==========")
    print(conf_matrix_cnn_lstm)

    # Lưu báo cáo vào file
    cnn_lstm_report_dir = "cnn_lstm_emotion_model"
    os.makedirs(cnn_lstm_report_dir, exist_ok=True)
    with open(os.path.join(cnn_lstm_report_dir, "classification_report.txt"), "w", encoding="utf-8") as f:
        f.write("========== CNN-LSTM Classification Report ==========\n")
        f.write(report_cnn_lstm)
        f.write("\n========== Additional Metrics ==========\n")
        f.write(f"Test Loss: {loss:.4f}\n")
        f.write(f"Test Accuracy: {test_accuracy_cnn_lstm:.4f}\n")
        f.write(f"Precision (Macro): {precision_macro_cnn_lstm:.4f}\n")
        f.write(f"Precision (Weighted): {precision_weighted_cnn_lstm:.4f}\n")
        f.write(f"Recall (Macro): {recall_macro_cnn_lstm:.4f}\n")
        f.write(f"Recall (Weighted): {recall_weighted_cnn_lstm:.4f}\n")
        f.write(f"F1-Score (Macro): {f1_macro_cnn_lstm:.4f}\n")
        f.write(f"F1-Score (Weighted): {f1_weighted_cnn_lstm:.4f}\n")
        f.write("\n========== Confusion Matrix ==========\n")
        f.write(np.array2string(conf_matrix_cnn_lstm))

    print("\n========== CNN-LSTM Classification Report saved to 'cnn_lstm_emotion_model/classification_report.txt' ==========")

    # Lưu mô hình CNN-LSTM
    model_cnn_lstm.save(os.path.join(cnn_lstm_report_dir, 'cnn_lstm_model.keras'))
    print(f"========== CNN-LSTM Model saved to '{cnn_lstm_report_dir}/cnn_lstm_model.keras' ==========")

    # ========== LƯU LABEL MAPPING VÀ VOCABULARY ==========
    # Lưu label_mapping và vocabulary cho RNN
    with open(os.path.join(rnn_report_dir, "label_mapping.json"), "w", encoding="utf-8") as f:
        json.dump(label_mapping, f, ensure_ascii=False, indent=4)

    with open(os.path.join(rnn_report_dir, "vocabulary.json"), "w", encoding="utf-8") as f:
        json.dump(data_manager.vocabulary.word2id, f, ensure_ascii=False, indent=4)

    # Lưu label_mapping và vocabulary cho CNN-LSTM
    # Giả sử label_mapping và vocabulary giống nhau, bạn có thể chỉ lưu một lần.
    # Nếu khác, hãy điều chỉnh tương ứng.

    print("========== Label Mapping and Vocabulary saved ==========")

    # ========== DEMO DỰ ĐOÁN 1 CÂU MỚI ==========

    custom_text = "Tôi rất vui khi sử dụng dịch vụ này!"

    # RNN (PyTorch)
    emotion_rnn = predict_emotion_rnn(
        model_rnn, custom_text, data_manager, label_mapping, device
    )
    print(f"Predicted Emotion (RNN): {emotion_rnn}")

    # CNN-LSTM (Keras)
    cnn_lstm_loaded = tf.keras.models.load_model(os.path.join(cnn_lstm_report_dir, 'cnn_lstm_model.keras'))
    emotion_cnn_lstm = predict_emotion_cnn_lstm(
        cnn_lstm_loaded, custom_text, data_manager, label_mapping
    )
    print(f"Predicted Emotion (CNN-LSTM): {emotion_cnn_lstm}")

    # Kiểm tra TF, GPU
    print("TF version:", tf.__version__)
    print("GPU devices:", tf.config.list_physical_devices("GPU"))
    # os.system("nvidia-smi")  # nếu muốn xem info GPU