{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.11.11","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"gpu","dataSources":[{"sourceId":8900,"databundleVersionId":862232,"sourceType":"competition"}],"dockerImageVersionId":31041,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"import os\nimport pandas as pd\nimport numpy as np\nimport torch\nimport torch.nn as nn\nimport torch.optim as optim\nfrom torch.utils.data import Dataset, DataLoader\nimport torchaudio\nimport matplotlib.pyplot as plt\nimport seaborn as sns\nfrom sklearn.metrics import confusion_matrix, accuracy_score, f1_score\nfrom sklearn.model_selection import train_test_split\nimport time\nimport librosa\nimport warnings\nimport math\nwarnings.filterwarnings('ignore')\n\n# Thiết lập seed để kết quả có thể tái tạo\ntorch.manual_seed(42)\nnp.random.seed(42)\n\n# Đường dẫn đến thư mục dữ liệu\nDATA_DIR = \"/kaggle/input/freesound-audio-tagging/audio_train\" # Thay đổi thành đường dẫn thư mục của bạn\nCSV_PATH = \"/kaggle/input/freesound-audio-tagging/train.csv\" # Thay đổi thành đường dẫn file CSV của bạn\n\n# Các tham số\nSAMPLE_RATE = 22050 # Tần số lấy mẫu chuẩn\nDURATION = 3 # Độ dài âm thanh (giây)\nN_MFCC = 40 # Số hệ số MFCC\nN_FFT = 2048 # Kích thước FFT\nHOP_LENGTH = 512 # Bước nhảy\nN_MELS = 128 # Số kênh Mel\nBATCH_SIZE = 1024\nEPOCHS = 20\nLEARNING_RATE = 0.0005\n\n# 1. ĐỌC VÀ XỬ LÝ DỮ LIỆU\n\nclass AudioDataset(Dataset):\n def __init__(self, annotations, audio_dir, transform=None, target_sr=SAMPLE_RATE, \n duration=DURATION, feature_type='mfcc'):\n \"\"\"\n Args:\n annotations (DataFrame): DataFrame chứa thông tin về các file âm thanh và nhãn\n audio_dir (string): Đường dẫn đến thư mục chứa file âm thanh\n transform (callable, optional): Optional transform to be applied on a sample\n target_sr (int): Sample rate mục tiêu\n duration (float): Độ dài mong muốn (giây)\n feature_type (str): Loại đặc trưng ('mfcc', 'melspec', 'logmel')\n \"\"\"\n self.annotations = annotations\n self.audio_dir = audio_dir\n self.transform = transform\n self.target_sr = target_sr\n self.duration = duration\n self.target_samples = int(target_sr * duration)\n self.feature_type = feature_type\n \n # Xác định số classes dựa trên dữ liệu\n self.num_classes = len(annotations['label'].unique())\n \n # Dictionary để map nhãn thành số nguyên\n unique_labels = sorted(annotations['label'].unique())\n self.label_to_idx = {label: i for i, label in enumerate(unique_labels)}\n self.idx_to_label = {i: label for i, label in enumerate(unique_labels)}\n \n def __len__(self):\n return len(self.annotations)\n \n def __getitem__(self, idx):\n audio_path = os.path.join(self.audio_dir, self.annotations.iloc[idx, 0])\n label = self.annotations.iloc[idx, 1]\n label_idx = self.label_to_idx[label]\n \n # Đọc file âm thanh\n try:\n waveform, sr = torchaudio.load(audio_path)\n \n # Chuyển đổi thành mono nếu có nhiều kênh\n if waveform.shape[0] > 1:\n waveform = torch.mean(waveform, dim=0, keepdim=True)\n \n # Resampling nếu cần\n if sr != self.target_sr:\n resampler = torchaudio.transforms.Resample(sr, self.target_sr)\n waveform = resampler(waveform)\n \n # Chuẩn hóa độ dài\n if waveform.shape[1] < self.target_samples:\n # Zero padding\n padding = self.target_samples - waveform.shape[1]\n waveform = torch.nn.functional.pad(waveform, (0, padding))\n elif waveform.shape[1] > self.target_samples:\n # Cắt bớt\n waveform = waveform[:, :self.target_samples]\n \n # Chuẩn hóa biên độ\n waveform = waveform / (torch.max(torch.abs(waveform)) + 1e-8)\n \n # Trích xuất đặc trưng dựa trên loại đã chọn\n if self.feature_type == 'mfcc':\n # MFCC\n mfcc_transform = torchaudio.transforms.MFCC(\n sample_rate=self.target_sr,\n n_mfcc=N_MFCC,\n melkwargs={'n_fft': N_FFT, 'hop_length': HOP_LENGTH, 'n_mels': N_MELS}\n )\n features = mfcc_transform(waveform) # Shape: [1, n_mfcc, time]\n \n elif self.feature_type == 'melspec':\n # Mel Spectrogram\n mel_transform = torchaudio.transforms.MelSpectrogram(\n sample_rate=self.target_sr,\n n_fft=N_FFT,\n hop_length=HOP_LENGTH,\n n_mels=N_MELS\n )\n features = mel_transform(waveform) # Shape: [1, n_mels, time]\n \n elif self.feature_type == 'logmel':\n # Log-Mel Spectrogram\n mel_transform = torchaudio.transforms.MelSpectrogram(\n sample_rate=self.target_sr,\n n_fft=N_FFT,\n hop_length=HOP_LENGTH,\n n_mels=N_MELS\n )\n features = mel_transform(waveform)\n features = torch.log(features + 1e-9) # Log scale\n \n # Áp dụng transform nếu được cung cấp\n if self.transform:\n features = self.transform(features)\n \n return features, label_idx\n \n except Exception as e:\n print(f\"Error loading {audio_path}: {e}\")\n # Trả về tensor rỗng trong trường hợp lỗi\n return torch.zeros(1, N_MFCC, int(self.target_samples / HOP_LENGTH) + 1), label_idx\n\n# 2. XÂY DỰNG CÁC MÔ HÌNH\n\n# CNN 1D\nclass CNN1D(nn.Module):\n def __init__(self, num_classes, input_size=N_MFCC):\n super(CNN1D, self).__init__()\n self.conv1 = nn.Conv1d(input_size, 64, kernel_size=3, stride=1, padding=1)\n self.bn1 = nn.BatchNorm1d(64)\n self.relu = nn.ReLU()\n self.pool1 = nn.MaxPool1d(kernel_size=2)\n \n self.conv2 = nn.Conv1d(64, 128, kernel_size=3, stride=1, padding=1)\n self.bn2 = nn.BatchNorm1d(128)\n self.pool2 = nn.MaxPool1d(kernel_size=2)\n \n self.conv3 = nn.Conv1d(128, 256, kernel_size=3, stride=1, padding=1)\n self.bn3 = nn.BatchNorm1d(256)\n self.pool3 = nn.MaxPool1d(kernel_size=2)\n \n # Tính toán kích thước đầu vào cho lớp fully connected\n self.fc_input_size = self._get_fc_input_size(input_size)\n \n self.fc1 = nn.Linear(self.fc_input_size, 512)\n self.dropout = nn.Dropout(0.5)\n self.fc2 = nn.Linear(512, num_classes)\n \n def _get_fc_input_size(self, input_size):\n # Tính toán kích thước đầu vào cho lớp fully connected\n with torch.no_grad():\n x = torch.zeros(1, input_size, 216) # Approximate time dimension\n x = self.pool3(self.bn3(self.relu(self.conv3(\n self.pool2(self.bn2(self.relu(self.conv2(\n self.pool1(self.bn1(self.relu(self.conv1(x))))))))))))\n return x.shape[1] * x.shape[2]\n \n def forward(self, x):\n # x shape: [batch, channels, time]\n x = self.pool1(self.relu(self.bn1(self.conv1(x))))\n x = self.pool2(self.relu(self.bn2(self.conv2(x))))\n x = self.pool3(self.relu(self.bn3(self.conv3(x))))\n \n # Reshape for fully connected layer\n x = x.view(x.size(0), -1)\n \n x = self.relu(self.fc1(x))\n x = self.dropout(x)\n x = self.fc2(x)\n return x\n\n# CNN 2D\nclass CNN2D(nn.Module):\n def __init__(self, num_classes, in_channels=1):\n super(CNN2D, self).__init__()\n self.conv1 = nn.Conv2d(in_channels, 32, kernel_size=3, stride=1, padding=1)\n self.bn1 = nn.BatchNorm2d(32)\n self.relu = nn.ReLU()\n self.pool1 = nn.MaxPool2d(kernel_size=2)\n \n self.conv2 = nn.Conv2d(32, 64, kernel_size=3, stride=1, padding=1)\n self.bn2 = nn.BatchNorm2d(64)\n self.pool2 = nn.MaxPool2d(kernel_size=2)\n \n self.conv3 = nn.Conv2d(64, 128, kernel_size=3, stride=1, padding=1)\n self.bn3 = nn.BatchNorm2d(128)\n self.pool3 = nn.MaxPool2d(kernel_size=2)\n \n # Tính toán kích thước đầu vào cho lớp fully connected\n self.fc_input_size = self._get_fc_input_size(in_channels)\n \n self.fc1 = nn.Linear(self.fc_input_size, 512)\n self.dropout = nn.Dropout(0.5)\n self.fc2 = nn.Linear(512, num_classes)\n \n def _get_fc_input_size(self, in_channels):\n # Tính toán kích thước đầu vào cho lớp fully connected\n with torch.no_grad():\n # Approximate dimensions for MFCC features\n x = torch.zeros(1, in_channels, N_MFCC, 216) # [batch, channels, height, width]\n x = self.pool3(self.bn3(self.relu(self.conv3(\n self.pool2(self.bn2(self.relu(self.conv2(\n self.pool1(self.bn1(self.relu(self.conv1(x))))))))))))\n return x.numel() // x.shape[0]\n \n def forward(self, x):\n # x shape: [batch, channels, height, width]\n x = self.pool1(self.relu(self.bn1(self.conv1(x))))\n x = self.pool2(self.relu(self.bn2(self.conv2(x))))\n x = self.pool3(self.relu(self.bn3(self.conv3(x))))\n \n # Reshape for fully connected layer\n x = x.view(x.size(0), -1)\n \n x = self.relu(self.fc1(x))\n x = self.dropout(x)\n x = self.fc2(x)\n return x\n\n# RNN\nclass RNN(nn.Module):\n def __init__(self, input_size, hidden_size, num_layers, num_classes):\n super(RNN, self).__init__()\n self.hidden_size = hidden_size\n self.num_layers = num_layers\n \n self.rnn = nn.RNN(input_size, hidden_size, num_layers, batch_first=True)\n self.fc = nn.Linear(hidden_size, num_classes)\n \n def forward(self, x):\n # x shape: [batch, time, features]\n # Set initial hidden states\n h0 = torch.zeros(self.num_layers, x.size(0), self.hidden_size).to(x.device)\n \n # Forward propagate RNN\n out, _ = self.rnn(x, h0)\n \n # Decode the hidden state of the last time step\n out = self.fc(out[:, -1, :])\n return out\n\n# LSTM\nclass LSTM(nn.Module):\n def __init__(self, input_size, hidden_size, num_layers, num_classes):\n super(LSTM, self).__init__()\n self.hidden_size = hidden_size\n self.num_layers = num_layers\n \n self.lstm = nn.LSTM(input_size, hidden_size, num_layers, batch_first=True)\n self.fc = nn.Linear(hidden_size, num_classes)\n \n def forward(self, x):\n # x shape: [batch, time, features]\n # Set initial hidden and cell states\n h0 = torch.zeros(self.num_layers, x.size(0), self.hidden_size).to(x.device)\n c0 = torch.zeros(self.num_layers, x.size(0), self.hidden_size).to(x.device)\n \n # Forward propagate LSTM\n out, _ = self.lstm(x, (h0, c0))\n \n # Decode the hidden state of the last time step\n out = self.fc(out[:, -1, :])\n return out\n\n# BiLSTM\nclass BiLSTM(nn.Module):\n def __init__(self, input_size, hidden_size, num_layers, num_classes):\n super(BiLSTM, self).__init__()\n self.hidden_size = hidden_size\n self.num_layers = num_layers\n \n self.lstm = nn.LSTM(input_size, hidden_size, num_layers, batch_first=True, bidirectional=True)\n self.fc = nn.Linear(hidden_size * 2, num_classes) # * 2 for bidirectional\n \n def forward(self, x):\n # x shape: [batch, time, features]\n # Set initial hidden and cell states\n h0 = torch.zeros(self.num_layers * 2, x.size(0), self.hidden_size).to(x.device) # * 2 for bidirectional\n c0 = torch.zeros(self.num_layers * 2, x.size(0), self.hidden_size).to(x.device)\n \n # Forward propagate LSTM\n out, _ = self.lstm(x, (h0, c0))\n \n # Decode the hidden state of the last time step\n out = self.fc(out[:, -1, :])\n return out\n\n# GRU\nclass GRU(nn.Module):\n def __init__(self, input_size, hidden_size, num_layers, num_classes):\n super(GRU, self).__init__()\n self.hidden_size = hidden_size\n self.num_layers = num_layers\n \n self.gru = nn.GRU(input_size, hidden_size, num_layers, batch_first=True)\n self.fc = nn.Linear(hidden_size, num_classes)\n \n def forward(self, x):\n # x shape: [batch, time, features]\n # Set initial hidden state\n h0 = torch.zeros(self.num_layers, x.size(0), self.hidden_size).to(x.device)\n \n # Forward propagate GRU\n out, _ = self.gru(x, h0)\n \n # Decode the hidden state of the last time step\n out = self.fc(out[:, -1, :])\n return out\n\nclass AudioTransformer(nn.Module):\n def __init__(self, input_size, num_classes, d_model=216, nhead=8, \n num_layers=3, dim_feedforward=512, dropout=0.1):\n super(AudioTransformer, self).__init__()\n \n self.d_model = d_model\n \n # Input projection layer\n self.input_projection = nn.Linear(input_size, d_model)\n \n # Positional encoding\n self.pos_encoder = PositionalEncoding(d_model, dropout)\n \n # Transformer encoder\n encoder_layers = nn.TransformerEncoderLayer(\n d_model=d_model,\n nhead=nhead,\n dim_feedforward=dim_feedforward,\n dropout=dropout,\n batch_first=True\n )\n self.transformer_encoder = nn.TransformerEncoder(encoder_layers, num_layers)\n \n # Output layer\n self.fc = nn.Linear(d_model, num_classes)\n \n def forward(self, x):\n # x shape: [batch, time, features]\n \n # Project input to d_model dimensions\n x = self.input_projection(x)\n \n # Add positional encoding\n x = self.pos_encoder(x)\n \n # Pass through transformer encoder\n x = self.transformer_encoder(x)\n \n # Global average pooling over time dimension\n x = torch.mean(x, dim=1)\n \n # Final classification\n x = self.fc(x)\n return x\n\nclass PositionalEncoding(nn.Module):\n def __init__(self, d_model, dropout=0.1, max_len=5000):\n super(PositionalEncoding, self).__init__()\n self.dropout = nn.Dropout(p=dropout)\n\n position = torch.arange(max_len).unsqueeze(1)\n div_term = torch.exp(torch.arange(0, d_model, 2) * (-math.log(10000.0) / d_model))\n pe = torch.zeros(max_len, 1, d_model)\n pe[:, 0, 0::2] = torch.sin(position * div_term)\n pe[:, 0, 1::2] = torch.cos(position * div_term)\n pe = pe.transpose(0, 1)\n self.register_buffer('pe', pe)\n\n def forward(self, x):\n x = x + self.pe[:, :x.size(1)]\n return self.dropout(x)\n\n# 3. HUẤN LUYỆN VÀ ĐÁNH GIÁ\n\ndef preprocess_data(feature_type='mfcc'):\n \"\"\"\n Đọc và tiền xử lý dữ liệu.\n \n Args:\n feature_type (str): Loại đặc trưng ('mfcc', 'melspec', 'logmel')\n \n Returns:\n tuple: (train_loader, val_loader, test_loader, num_classes)\n \"\"\"\n # Đọc file CSV\n df = pd.read_csv(CSV_PATH)\n \n # Đổi tên cột nếu cần thiết\n if 'fname' in df.columns and 'label' in df.columns:\n pass\n else:\n # Đặt tên cột chính xác từ dữ liệu mô tả\n df.columns = ['fname', 'label', 'manually_verified']\n \n # Lấy những dòng có manually_verified = 1 nếu cần\n # df = df[df['manually_verified'] == 1]\n \n # Chia dữ liệu thành train, validation và test\n train_df, temp_df = train_test_split(df, test_size=0.3, random_state=42, stratify=df['label'])\n val_df, test_df = train_test_split(temp_df, test_size=0.5, random_state=42, stratify=temp_df['label'])\n \n print(f\"Train set size: {len(train_df)}\")\n print(f\"Validation set size: {len(val_df)}\")\n print(f\"Test set size: {len(test_df)}\")\n \n # Khởi tạo datasets\n train_dataset = AudioDataset(train_df, DATA_DIR, feature_type=feature_type)\n val_dataset = AudioDataset(val_df, DATA_DIR, feature_type=feature_type)\n test_dataset = AudioDataset(test_df, DATA_DIR, feature_type=feature_type)\n \n # Tạo data loaders\n train_loader = DataLoader(train_dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=4)\n val_loader = DataLoader(val_dataset, batch_size=BATCH_SIZE, shuffle=False, num_workers=4)\n test_loader = DataLoader(test_dataset, batch_size=BATCH_SIZE, shuffle=False, num_workers=4)\n \n num_classes = train_dataset.num_classes\n \n return train_loader, val_loader, test_loader, num_classes, train_dataset.idx_to_label\n\ndef train_model(model, train_loader, val_loader, device, model_name, feature_type):\n \"\"\"\n Huấn luyện và đánh giá mô hình.\n \n Args:\n model: PyTorch model\n train_loader: DataLoader cho tập huấn luyện\n val_loader: DataLoader cho tập validation\n device: Thiết bị để huấn luyện (CPU hoặc GPU)\n model_name: Tên của mô hình\n feature_type: Loại đặc trưng được sử dụng\n \n Returns:\n dict: Lịch sử huấn luyện và thời gian đào tạo\n \"\"\"\n criterion = nn.CrossEntropyLoss()\n optimizer = optim.Adam(model.parameters(), lr=LEARNING_RATE)\n \n # Learning rate scheduler\n scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', factor=0.5, patience=3, verbose=True)\n \n # Tracking metrics\n history = {\n 'train_loss': [],\n 'val_loss': [],\n 'train_acc': [],\n 'val_acc': []\n }\n \n start_time = time.time()\n best_val_loss = float('inf')\n \n # Training loop\n for epoch in range(EPOCHS):\n # Training\n model.train()\n train_loss = 0.0\n train_correct = 0\n train_total = 0\n \n for features, labels in train_loader:\n features, labels = features.to(device), labels.to(device)\n \n # Đảm bảo đầu vào phù hợp với kiến trúc mô hình\n if isinstance(model, CNN1D):\n # CNN1D nhận đầu vào dạng [batch, features, time]\n if features.dim() == 4: # [batch, channels, height, time]\n features = features.squeeze(1) # Remove channel dimension\n elif isinstance(model, CNN2D):\n # CNN2D nhận đầu vào dạng [batch, channels, height, width]\n if features.dim() == 3: # [batch, height, time]\n features = features.unsqueeze(1) # Add channel dimension\n elif isinstance(model, (RNN, LSTM, BiLSTM, GRU)):\n # RNN models nhận đầu vào dạng [batch, time, features]\n if features.dim() == 3: # [batch, features, time]\n features = features.transpose(1, 2)\n elif features.dim() == 4: # [batch, channels, height, time]\n b, c, h, w = features.shape\n features = features.reshape(b, w, c*h) # [batch, time, channels*height]\n \n optimizer.zero_grad()\n outputs = model(features)\n loss = criterion(outputs, labels)\n loss.backward()\n optimizer.step()\n \n train_loss += loss.item()\n _, predicted = torch.max(outputs.data, 1)\n train_total += labels.size(0)\n train_correct += (predicted == labels).sum().item()\n \n train_loss = train_loss / len(train_loader)\n train_acc = train_correct / train_total\n \n # Validation\n model.eval()\n val_loss = 0.0\n val_correct = 0\n val_total = 0\n \n with torch.no_grad():\n for features, labels in val_loader:\n features, labels = features.to(device), labels.to(device)\n \n # Đảm bảo đầu vào phù hợp với kiến trúc mô hình\n if isinstance(model, CNN1D):\n if features.dim() == 4:\n features = features.squeeze(1)\n elif isinstance(model, CNN2D):\n if features.dim() == 3:\n features = features.unsqueeze(1)\n elif isinstance(model, (RNN, LSTM, BiLSTM, GRU)):\n if features.dim() == 3:\n features = features.transpose(1, 2)\n elif features.dim() == 4:\n b, c, h, w = features.shape\n features = features.reshape(b, w, c*h)\n \n outputs = model(features)\n loss = criterion(outputs, labels)\n \n val_loss += loss.item()\n _, predicted = torch.max(outputs.data, 1)\n val_total += labels.size(0)\n val_correct += (predicted == labels).sum().item()\n \n val_loss = val_loss / len(val_loader)\n val_acc = val_correct / val_total\n \n # Learning rate scheduler step\n scheduler.step(val_loss)\n \n # Save best model\n if val_loss < best_val_loss:\n best_val_loss = val_loss\n torch.save(model.state_dict(), f\"{model_name}_{feature_type}_best.pth\")\n \n # Update history\n history['train_loss'].append(train_loss)\n history['val_loss'].append(val_loss)\n history['train_acc'].append(train_acc)\n history['val_acc'].append(val_acc)\n \n print(f\"Epoch {epoch+1}/{EPOCHS}, \"\n f\"Train Loss: {train_loss:.4f}, Train Acc: {train_acc:.4f}, \"\n f\"Val Loss: {val_loss:.4f}, Val Acc: {val_acc:.4f}\")\n \n end_time = time.time()\n train_time = end_time - start_time\n print(f\"Training time: {train_time:.2f} seconds\")\n \n # Plot learning curves\n plt.figure(figsize=(12, 5))\n \n plt.subplot(1, 2, 1)\n plt.plot(history['train_loss'], label='Train')\n plt.plot(history['val_loss'], label='Validation')\n plt.title(f'{model_name} - Loss')\n plt.xlabel('Epoch')\n plt.ylabel('Loss')\n plt.legend()\n \n plt.subplot(1, 2, 2)\n plt.plot(history['train_acc'], label='Train')\n plt.plot(history['val_acc'], label='Validation')\n plt.title(f'{model_name} - Accuracy')\n plt.xlabel('Epoch')\n plt.ylabel('Accuracy')\n plt.legend()\n \n plt.tight_layout()\n plt.savefig(f\"{model_name}_{feature_type}_learning_curves.png\")\n plt.close()\n \n return history, train_time\n\ndef evaluate_model(model, test_loader, device, idx_to_label, model_name, feature_type):\n \"\"\"\n Đánh giá mô hình trên tập test.\n \n Args:\n model: PyTorch model\n test_loader: DataLoader cho tập test\n device: Thiết bị để đánh giá (CPU hoặc GPU)\n idx_to_label: Dictionary để chuyển đổi index thành nhãn\n model_name: Tên của mô hình\n feature_type: Loại đặc trưng được sử dụng\n \n Returns:\n tuple: (accuracy, f1, confusion matrix)\n \"\"\"\n model.eval()\n all_preds = []\n all_labels = []\n \n start_time = time.time()\n \n with torch.no_grad():\n for features, labels in test_loader:\n features, labels = features.to(device), labels.to(device)\n \n # Đảm bảo đầu vào phù hợp với kiến trúc mô hình\n if isinstance(model, CNN1D):\n if features.dim() == 4:\n features = features.squeeze(1)\n elif isinstance(model, CNN2D):\n if features.dim() == 3:\n features = features.unsqueeze(1)\n elif isinstance(model, (RNN, LSTM, BiLSTM, GRU)):\n if features.dim() == 3:\n features = features.transpose(1, 2)\n elif features.dim() == 4:\n b, c, h, w = features.shape\n features = features.reshape(b, w, c*h)\n \n outputs = model(features)\n _, preds = torch.max(outputs, 1)\n \n all_preds.extend(preds.cpu().numpy())\n all_labels.extend(labels.cpu().numpy())\n \n end_time = time.time()\n test_time = end_time - start_time\n \n # Calculate metrics\n accuracy = accuracy_score(all_labels, all_preds)\n f1 = f1_score(all_labels, all_preds, average='weighted') # Weighted average for multiclass\n cm = confusion_matrix(all_labels, all_preds)\n \n print(f\"\\nTest Results for {model_name}:\")\n print(f\"Accuracy: {accuracy:.4f}\")\n print(f\"F1 Score: {f1:.4f}\")\n print(f\"Test time: {test_time:.2f} seconds\")\n \n # Plot confusion matrix\n plt.figure(figsize=(10, 8))\n sns.heatmap(cm, annot=True, fmt='d', cmap='Blues',\n xticklabels=[idx_to_label[i] for i in range(len(idx_to_label))],\n yticklabels=[idx_to_label[i] for i in range(len(idx_to_label))])\n plt.xlabel('Predicted')\n plt.ylabel('True')\n plt.title(f'Confusion Matrix - {model_name}')\n plt.tight_layout()\n plt.savefig(f\"{model_name}_{feature_type}_confusion_matrix.png\")\n plt.close()\n \n return accuracy, f1, cm, test_time\n\ndef main():\n device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n print(f\"Using device: {device}\")\n \n # Định nghĩa các loại đặc trưng cần thử nghiệm\n feature_types = ['mfcc']\n \n # Lưu trữ kết quả\n results = {\n 'model_name': [],\n 'feature_type': [],\n 'accuracy': [],\n 'f1_score': [],\n 'train_time': [],\n 'test_time': []\n }\n\n # Thử nghiệm với các loại đặc trưng khác nhau\n for feature_type in feature_types:\n print(f\"\\nProcessing feature type: {feature_type}\")\n \n # Load và tiền xử lý dữ liệu\n train_loader, val_loader, test_loader, num_classes, idx_to_label = preprocess_data(feature_type)\n \n # Định nghĩa các mô hình cần thử nghiệm\n models = {\n 'CNN1D': CNN1D(num_classes=num_classes),\n 'CNN2D': CNN2D(num_classes=num_classes),\n 'RNN': RNN(input_size=N_MFCC, hidden_size=128, num_layers=2, num_classes=num_classes),\n 'LSTM': LSTM(input_size=N_MFCC, hidden_size=128, num_layers=2, num_classes=num_classes),\n 'BiLSTM': BiLSTM(input_size=N_MFCC, hidden_size=128, num_layers=2, num_classes=num_classes),\n 'GRU': GRU(input_size=N_MFCC, hidden_size=128, num_layers=2, num_classes=num_classes),\n 'Transformer': AudioTransformer(\n input_size=N_MFCC,\n num_classes=num_classes,\n d_model=216,\n nhead=8,\n num_layers=3,\n dim_feedforward=512,\n dropout=0.1\n )\n }\n \n # Huấn luyện và đánh giá từng mô hình\n for model_name, model in models.items():\n print(f\"\\nTraining {model_name} with {feature_type} features...\")\n model = model.to(device)\n \n # Huấn luyện mô hình\n history, train_time = train_model(model, train_loader, val_loader, device, model_name, feature_type)\n \n # Load best model weights\n model.load_state_dict(torch.load(f\"{model_name}_{feature_type}_best.pth\"))\n \n # Đánh giá mô hình\n accuracy, f1, cm, test_time = evaluate_model(model, test_loader, device, idx_to_label, model_name, feature_type)\n \n # Lưu kết quả\n results['model_name'].append(model_name)\n results['feature_type'].append(feature_type)\n results['accuracy'].append(accuracy)\n results['f1_score'].append(f1)\n results['train_time'].append(train_time)\n results['test_time'].append(test_time)\n \n # Tạo DataFrame từ kết quả và lưu vào file CSV\n results_df = pd.DataFrame(results)\n results_df.to_csv('model_comparison_results.csv', index=False)\n \n # Hiển thị bảng kết quả\n print(\"\\nFinal Results:\")\n print(results_df)\n \n # Vẽ biểu đồ so sánh\n plt.figure(figsize=(15, 6))\n \n plt.subplot(1, 2, 1)\n sns.barplot(data=results_df, x='model_name', y='accuracy', hue='feature_type')\n plt.title('Model Accuracy Comparison')\n plt.xticks(rotation=45)\n \n plt.subplot(1, 2, 2)\n sns.barplot(data=results_df, x='model_name', y='f1_score', hue='feature_type')\n plt.title('Model F1 Score Comparison')\n plt.xticks(rotation=45)\n \n plt.tight_layout()\n plt.savefig('model_comparison.png')\n plt.close()\n\nif __name__ == \"__main__\":\n main()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-05-22T06:29:50.440032Z","iopub.execute_input":"2025-05-22T06:29:50.440543Z","iopub.status.idle":"2025-05-22T06:31:11.565387Z","shell.execute_reply.started":"2025-05-22T06:29:50.440519Z","shell.execute_reply":"2025-05-22T06:31:11.564181Z"}},"outputs":[{"name":"stdout","text":"Using device: cuda\n\nProcessing feature type: mfcc\nTrain set size: 6631\nValidation set size: 1421\nTest set size: 1421\n\nTraining CNN1D with mfcc features...\n","output_type":"stream"},{"traceback":["\u001b[0;31m---------------------------------------------------------------------------\u001b[0m","\u001b[0;31mRuntimeError\u001b[0m Traceback (most recent call last)","\u001b[0;32m/tmp/ipykernel_35/237653695.py\u001b[0m in \u001b[0;36m\u001b[0;34m()\u001b[0m\n\u001b[1;32m 756\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 757\u001b[0m \u001b[0;32mif\u001b[0m \u001b[0m__name__\u001b[0m \u001b[0;34m==\u001b[0m \u001b[0;34m\"__main__\"\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 758\u001b[0;31m \u001b[0mmain\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m","\u001b[0;32m/tmp/ipykernel_35/237653695.py\u001b[0m in \u001b[0;36mmain\u001b[0;34m()\u001b[0m\n\u001b[1;32m 714\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 715\u001b[0m \u001b[0;31m# Huấn luyện mô hình\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 716\u001b[0;31m \u001b[0mhistory\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mtrain_time\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mtrain_model\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mmodel\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mtrain_loader\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mval_loader\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mdevice\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mmodel_name\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mfeature_type\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 717\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 718\u001b[0m \u001b[0;31m# Load best model weights\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n","\u001b[0;32m/tmp/ipykernel_35/237653695.py\u001b[0m in \u001b[0;36mtrain_model\u001b[0;34m(model, train_loader, val_loader, device, model_name, feature_type)\u001b[0m\n\u001b[1;32m 498\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 499\u001b[0m \u001b[0moptimizer\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mzero_grad\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 500\u001b[0;31m \u001b[0moutputs\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mmodel\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mfeatures\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 501\u001b[0m \u001b[0mloss\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mcriterion\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0moutputs\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mlabels\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 502\u001b[0m \u001b[0mloss\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mbackward\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n","\u001b[0;32m/usr/local/lib/python3.11/dist-packages/torch/nn/modules/module.py\u001b[0m in \u001b[0;36m_wrapped_call_impl\u001b[0;34m(self, *args, **kwargs)\u001b[0m\n\u001b[1;32m 1737\u001b[0m \u001b[0;32mreturn\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0m_compiled_call_impl\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m*\u001b[0m\u001b[0margs\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m \u001b[0;31m# type: ignore[misc]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 1738\u001b[0m \u001b[0;32melse\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m-> 1739\u001b[0;31m \u001b[0;32mreturn\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0m_call_impl\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m*\u001b[0m\u001b[0margs\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 1740\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 1741\u001b[0m \u001b[0;31m# torchrec tests the code consistency with the following code\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n","\u001b[0;32m/usr/local/lib/python3.11/dist-packages/torch/nn/modules/module.py\u001b[0m in \u001b[0;36m_call_impl\u001b[0;34m(self, *args, **kwargs)\u001b[0m\n\u001b[1;32m 1748\u001b[0m \u001b[0;32mor\u001b[0m \u001b[0m_global_backward_pre_hooks\u001b[0m \u001b[0;32mor\u001b[0m \u001b[0m_global_backward_hooks\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 1749\u001b[0m or _global_forward_hooks or _global_forward_pre_hooks):\n\u001b[0;32m-> 1750\u001b[0;31m \u001b[0;32mreturn\u001b[0m \u001b[0mforward_call\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m*\u001b[0m\u001b[0margs\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 1751\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 1752\u001b[0m \u001b[0mresult\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0;32mNone\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n","\u001b[0;32m/tmp/ipykernel_35/237653695.py\u001b[0m in \u001b[0;36mforward\u001b[0;34m(self, x)\u001b[0m\n\u001b[1;32m 185\u001b[0m \u001b[0mx\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mx\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mview\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mx\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0msize\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;36m0\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m-\u001b[0m\u001b[0;36m1\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 186\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 187\u001b[0;31m \u001b[0mx\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mrelu\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mfc1\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mx\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 188\u001b[0m \u001b[0mx\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mdropout\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mx\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 189\u001b[0m \u001b[0mx\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mfc2\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mx\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n","\u001b[0;32m/usr/local/lib/python3.11/dist-packages/torch/nn/modules/module.py\u001b[0m in \u001b[0;36m_wrapped_call_impl\u001b[0;34m(self, *args, **kwargs)\u001b[0m\n\u001b[1;32m 1737\u001b[0m \u001b[0;32mreturn\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0m_compiled_call_impl\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m*\u001b[0m\u001b[0margs\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m \u001b[0;31m# type: ignore[misc]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 1738\u001b[0m \u001b[0;32melse\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m-> 1739\u001b[0;31m \u001b[0;32mreturn\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0m_call_impl\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m*\u001b[0m\u001b[0margs\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 1740\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 1741\u001b[0m \u001b[0;31m# torchrec tests the code consistency with the following code\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n","\u001b[0;32m/usr/local/lib/python3.11/dist-packages/torch/nn/modules/module.py\u001b[0m in \u001b[0;36m_call_impl\u001b[0;34m(self, *args, **kwargs)\u001b[0m\n\u001b[1;32m 1748\u001b[0m \u001b[0;32mor\u001b[0m \u001b[0m_global_backward_pre_hooks\u001b[0m \u001b[0;32mor\u001b[0m \u001b[0m_global_backward_hooks\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 1749\u001b[0m or _global_forward_hooks or _global_forward_pre_hooks):\n\u001b[0;32m-> 1750\u001b[0;31m \u001b[0;32mreturn\u001b[0m \u001b[0mforward_call\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m*\u001b[0m\u001b[0margs\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 1751\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 1752\u001b[0m \u001b[0mresult\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0;32mNone\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n","\u001b[0;32m/usr/local/lib/python3.11/dist-packages/torch/nn/modules/linear.py\u001b[0m in \u001b[0;36mforward\u001b[0;34m(self, input)\u001b[0m\n\u001b[1;32m 123\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 124\u001b[0m \u001b[0;32mdef\u001b[0m \u001b[0mforward\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0minput\u001b[0m\u001b[0;34m:\u001b[0m \u001b[0mTensor\u001b[0m\u001b[0;34m)\u001b[0m \u001b[0;34m->\u001b[0m \u001b[0mTensor\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 125\u001b[0;31m \u001b[0;32mreturn\u001b[0m \u001b[0mF\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mlinear\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0minput\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mweight\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mbias\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 126\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 127\u001b[0m \u001b[0;32mdef\u001b[0m \u001b[0mextra_repr\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m)\u001b[0m \u001b[0;34m->\u001b[0m \u001b[0mstr\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n","\u001b[0;31mRuntimeError\u001b[0m: mat1 and mat2 shapes cannot be multiplied (1024x2560 and 6912x512)"],"ename":"RuntimeError","evalue":"mat1 and mat2 shapes cannot be multiplied (1024x2560 and 6912x512)","output_type":"error"}],"execution_count":1},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null}]}