{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.10.12","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"nvidiaTeslaT4","dataSources":[{"sourceId":91844,"databundleVersionId":11361821,"sourceType":"competition"},{"sourceId":11053663,"sourceType":"datasetVersion","datasetId":6886569}],"dockerImageVersionId":30919,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"# **BirdCLEF 2025 Training Notebook**\n\nThis is a baseline training pipeline for BirdCLEF 2025 using EfficientNetB0 with PyTorch and Timm(for pretrained EffNet). You can check inference and preprocessing notebooks in the following links: \n\n- [EfficientNet B0 Pytorch [Inference] | BirdCLEF'25](https://www.kaggle.com/code/kadircandrisolu/efficientnet-b0-pytorch-inference-birdclef-25)\n\n  \n- [Transforming Audio-to-Mel Spec. | BirdCLEF'25](https://www.kaggle.com/code/kadircandrisolu/transforming-audio-to-mel-spec-birdclef-25)  \n\nNote that by default this notebook is in Debug Mode, so it will only train the model with 2 epochs, but the [weight](https://www.kaggle.com/datasets/kadircandrisolu/birdclef25-effnetb0-starter-weight) I used in the inference notebook was obtained after 10 epochs of training.\n\n**Features**\n* Implement with Pytorch and Timm\n* Flexible audio processing with both pre-computed and on-the-fly mel spectrograms\n* Stratified 5-fold cross-validation with ensemble capability\n* Mixup training for improved generalization\n* Spectrogram augmentations (time/frequency masking, brightness adjustment)\n* AdamW optimizer with Cosine Annealing LR scheduling\n* Debug mode for quick experimentation with smaller datasets\n\n**Pre-computed Spectrograms**\nFor faster training, you can use pre-computed mel spectrograms from [this dataset](https://www.kaggle.com/datasets/kadircandrisolu/birdclef25-mel-spectrograms) by setting `LOAD_DATA = True`","metadata":{}},{"cell_type":"markdown","source":"## Libraries","metadata":{}},{"cell_type":"code","source":"import os\nimport logging\nimport random\nimport gc\nimport time\nimport cv2\nimport math\nimport warnings\nfrom pathlib import Path\n\nimport numpy as np\nimport pandas as pd\nfrom sklearn.model_selection import StratifiedKFold\nfrom sklearn.metrics import roc_auc_score\nimport librosa\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torch.optim as optim\nfrom torch.optim import lr_scheduler\nfrom torch.utils.data import Dataset, DataLoader\n\nimport matplotlib.pyplot as plt\nimport seaborn as sns\nfrom tqdm.auto import tqdm\n\nimport timm\n\nwarnings.filterwarnings(\"ignore\")\nlogging.basicConfig(level=logging.ERROR)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-05-14T10:52:27.928634Z","iopub.execute_input":"2025-05-14T10:52:27.928961Z","iopub.status.idle":"2025-05-14T10:52:27.934435Z","shell.execute_reply.started":"2025-05-14T10:52:27.928936Z","shell.execute_reply":"2025-05-14T10:52:27.933703Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Configuration","metadata":{}},{"cell_type":"code","source":"class CFG:\n    \n    seed = 42   # 随机种子，保证实验可复现性\n    debug = False    # 调试模式标志(True时会减少数据量和训练轮次)\n    apex = False   # 是否使用NVIDIA的混合精度训练\n    print_freq = 100   # 训练过程中打印信息的频率(每多少batch)\n    num_workers = 2    # 数据加载的线程数\n    \n    OUTPUT_DIR = '/kaggle/working/'    # 输出目录(Kaggle工作目录)\n\n    # 数据路径(BirdCLEF 2025比赛数据)\n    train_datadir = '/kaggle/input/birdclef-2025/train_audio'  # 训练音频文件\n    train_csv = '/kaggle/input/birdclef-2025/train.csv'   # 训练标签\n    test_soundscapes = '/kaggle/input/birdclef-2025/test_soundscapes'   # 测试音频\n    submission_csv = '/kaggle/input/birdclef-2025/sample_submission.csv'  # 提交样例\n    taxonomy_csv = '/kaggle/input/birdclef-2025/taxonomy.csv'   # 鸟类分类信息\n\n    # 预处理的频谱图数据\n    spectrogram_npy = '/kaggle/input/birdclef25-mel-spectrograms/birdclef2025_melspec_5sec_256_256.npy'\n \n    model_name = 'efficientnet_b0'    # 使用的模型架构(EfficientNet B0)\n    pretrained = True   # 是否使用预训练权重\n    in_channels = 1   # 输入通道数(单通道灰度频谱图)\n\n    LOAD_DATA = True    # 是否加载预处理数据(False时从原始音频处理)\n    FS = 32000               # 音频采样率(32kHz)\n    TARGET_DURATION = 10.0    # 每段音频的目标时长(秒)   TARGET_DURATION\n    TARGET_SHAPE = (256, 256)  # 频谱图的目标尺寸(高度,宽度)\n\n    # 频谱图生成参数\n    N_FFT = 1024             # FFT窗口大小\n    HOP_LENGTH = 512         # 帧移\n    N_MELS = 128             # 梅尔带数量\n    FMIN = 50                # 最小频率(Hz)\n    FMAX = 14000             # 最大频率(Hz)\n    \n    device = 'cuda' if torch.cuda.is_available() else 'cpu'\n    epochs = 10    # 训练总轮次\n    batch_size = 32    # 批量大小\n    criterion = 'BCEWithLogitsLoss'  # 损失函数(二分类交叉熵+Logits)\n\n    # 交叉验证设置\n    #n_fold = 5   # 总折数\n    n_fold = 7   # 总折数\n    selected_folds = [0, 1, 2, 3, 4, 5 ,6 ]    # 实际使用的折(调试时可减少)\n\n    # 优化器设置\n    optimizer = 'AdamW'   # 优化器类型\n    lr = 5e-4   # 初始学习率\n    weight_decay = 1e-5    # 权重衰减(L2正则化)\n  \n    scheduler = 'CosineAnnealingLR'   # 余弦退火学习率\n    min_lr = 1e-6   # 最小学习率\n    T_max = epochs   # 余弦周期长度(这里设为总epoch数)\n\n    aug_prob = 0.5   # 数据增强应用概率\n    mixup_alpha = 0.5    # Mixup增强的alpha参数\n    \n    def update_debug_settings(self):\n        if self.debug:\n            self.epochs = 2   # 调试时减少训练轮次\n            self.selected_folds = [0]   # 只使用第一折数据\n\ncfg = CFG()  # 实例化配置对象","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-05-14T10:52:28.948745Z","iopub.execute_input":"2025-05-14T10:52:28.948977Z","iopub.status.idle":"2025-05-14T10:52:28.956033Z","shell.execute_reply.started":"2025-05-14T10:52:28.948958Z","shell.execute_reply":"2025-05-14T10:52:28.955194Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Utilities","metadata":{}},{"cell_type":"code","source":"#设置随机种子的函数，默认使用著名的\"42\"作为种子值\ndef set_seed(seed=42):\n    \"\"\"\n    Set seed for reproducibility\n    \"\"\"\n    random.seed(seed)   # 设置Python内置随机模块种子\n    os.environ[\"PYTHONHASHSEED\"] = str(seed)    # 设置Python哈希种子环境变量(防止哈希随机化)\n    np.random.seed(seed)   # 设置NumPy随机种子\n    torch.manual_seed(seed)    # 设置PyTorch CPU随机种子\n    torch.cuda.manual_seed(seed)   # 设置当前GPU随机种子\n    torch.cuda.manual_seed_all(seed)   # 设置所有GPU随机种子(多GPU环境)\n    torch.backends.cudnn.deterministic = True   # 启用CuDNN确定性算法(确保可重复性，但可能降低性能)\n    torch.backends.cudnn.benchmark = False   # 禁用CuDNN自动优化基准(防止算法选择导致不一致)\n\n# 使用配置中的种子值初始化随机状态\nset_seed(cfg.seed)  # cfg.seed在前文定义为42","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-05-14T10:52:29.347298Z","iopub.execute_input":"2025-05-14T10:52:29.347507Z","iopub.status.idle":"2025-05-14T10:52:29.353659Z","shell.execute_reply.started":"2025-05-14T10:52:29.347489Z","shell.execute_reply":"2025-05-14T10:52:29.353026Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Pre-processing\nThese functions handle the transformation of audio files to mel spectrograms for model input, with flexibility controlled by the `LOAD_DATA` parameter. The process involves either loading pre-computed spectrograms from this [dataset](https://www.kaggle.com/datasets/kadircandrisolu/birdclef25-mel-spectrograms) (when `LOAD_DATA=True`) or dynamically generating them (when `LOAD_DATA=False`), transforming audio data into spectrogram representations, and preparing it for the neural network.","metadata":{}},{"cell_type":"code","source":"def audio2melspec(audio_data, cfg):\n    \"\"\"Convert audio data to mel spectrogram\"\"\"\n    # 处理NaN值：用均值填充缺失值\n    if np.isnan(audio_data).any():\n        mean_signal = np.nanmean(audio_data)\n        audio_data = np.nan_to_num(audio_data, nan=mean_signal)\n\n    # 生成梅尔频谱图(功率谱)\n    mel_spec = librosa.feature.melspectrogram(\n        y=audio_data,\n        sr=cfg.FS,\n        n_fft=cfg.N_FFT,\n        hop_length=cfg.HOP_LENGTH,\n        n_mels=cfg.N_MELS,\n        fmin=cfg.FMIN,\n        fmax=cfg.FMAX,\n        power=2.0\n    )\n\n    # 转换为分贝单位(对数变换)\n    mel_spec_db = librosa.power_to_db(mel_spec, ref=np.max)\n    # 归一化到[0,1]范围(防止除以0加1e-8)\n    mel_spec_norm = (mel_spec_db - mel_spec_db.min()) / (mel_spec_db.max() - mel_spec_db.min() + 1e-8)\n    \n    return mel_spec_norm\n\ndef process_audio_file(audio_path, cfg):\n    \"\"\"Process a single audio file to get the mel spectrogram\"\"\"\n    try:\n        # 加载音频文件(自动重采样到目标采样率)\n        audio_data, _ = librosa.load(audio_path, sr=cfg.FS)\n\n        # 计算目标样本数 = 时长(秒) × 采样率\n        target_samples = int(cfg.TARGET_DURATION * cfg.FS)\n\n        # 处理短音频：通过重复填充到目标长度\n        if len(audio_data) < target_samples:\n            n_copy = math.ceil(target_samples / len(audio_data))\n            if n_copy > 1:\n                audio_data = np.concatenate([audio_data] * n_copy)\n\n        # 截取中心10秒片段(避免开头/结尾静音)\n        # Extract center 10 seconds\n        start_idx = max(0, int(len(audio_data) / 2  - target_samples / 2))\n        end_idx = min(len(audio_data), start_idx + target_samples)\n        center_audio = audio_data[start_idx:end_idx]\n\n        # 如果截取的片段不足10秒，则进行填充\n        if len(center_audio) < target_samples:\n            center_audio = np.pad(center_audio, \n                                 (0, target_samples - len(center_audio)), \n                                 mode='constant')\n\n        # 生成梅尔频谱图\n        mel_spec = audio2melspec(center_audio, cfg)\n        \n        if mel_spec.shape != cfg.TARGET_SHAPE:\n            mel_spec = cv2.resize(mel_spec, cfg.TARGET_SHAPE, interpolation=cv2.INTER_LINEAR)\n\n        return mel_spec.astype(np.float32)\n        \n    except Exception as e:\n        print(f\"Error processing {audio_path}: {e}\")\n        return None\n\ndef generate_spectrograms(df, cfg):\n    \"\"\"Generate spectrograms from audio files\"\"\"\n    print(\"Generating mel spectrograms from audio files...\")\n    start_time = time.time()\n\n    all_bird_data = {}   # 存储成功处理的频谱图\n    errors = []\n\n    for i, row in tqdm(df.iterrows(), total=len(df)):\n        # 调试模式只处理前1000个样本\n        if cfg.debug and i >= 1000:\n            break\n        \n        try:\n            samplename = row['samplename']   # 音频唯一标识\n            filepath = row['filepath']  # 音频文件路径\n            \n            mel_spec = process_audio_file(filepath, cfg)\n            \n            # 处理单个音频文件\n            if mel_spec is not None:\n                all_bird_data[samplename] = mel_spec\n            \n        except Exception as e:\n            print(f\"Error processing {row.filepath}: {e}\")\n            errors.append((row.filepath, str(e)))\n\n    end_time = time.time()\n    print(f\"Processing completed in {end_time - start_time:.2f} seconds\")\n    print(f\"Successfully processed {len(all_bird_data)} files out of {len(df)}\")\n    print(f\"Failed to process {len(errors)} files\")\n    \n    return all_bird_data","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-05-14T10:52:29.693015Z","iopub.execute_input":"2025-05-14T10:52:29.693253Z","iopub.status.idle":"2025-05-14T10:52:29.702602Z","shell.execute_reply.started":"2025-05-14T10:52:29.693233Z","shell.execute_reply":"2025-05-14T10:52:29.701904Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Dataset Preparation and Data Augmentations\nWe'll convert audio to mel spectrograms and apply random augmentations with 50% probability each - including time stretching, pitch shifting, and volume adjustments. This randomized approach creates diverse training samples from the same audio files","metadata":{}},{"cell_type":"code","source":"class BirdCLEFDatasetFromNPY(Dataset):\n    def __init__(self, df, cfg, spectrograms=None, mode=\"train\"):\n        self.df = df\n        self.cfg = cfg\n        self.mode = mode\n\n        self.spectrograms = spectrograms\n        \n        taxonomy_df = pd.read_csv(self.cfg.taxonomy_csv)\n        self.species_ids = taxonomy_df['primary_label'].tolist()\n        self.num_classes = len(self.species_ids)\n        self.label_to_idx = {label: idx for idx, label in enumerate(self.species_ids)}\n\n        if 'filepath' not in self.df.columns:\n            self.df['filepath'] = self.cfg.train_datadir + '/' + self.df.filename\n        \n        if 'samplename' not in self.df.columns:\n            self.df['samplename'] = self.df.filename.map(lambda x: x.split('/')[0] + '-' + x.split('/')[-1].split('.')[0])\n\n        sample_names = set(self.df['samplename'])\n        if self.spectrograms:\n            found_samples = sum(1 for name in sample_names if name in self.spectrograms)\n            print(f\"Found {found_samples} matching spectrograms for {mode} dataset out of {len(self.df)} samples\")\n        \n        if cfg.debug:\n            self.df = self.df.sample(min(1000, len(self.df)), random_state=cfg.seed).reset_index(drop=True)\n    \n    def __len__(self):\n        return len(self.df)\n    \n    def __getitem__(self, idx):\n        row = self.df.iloc[idx]\n        samplename = row['samplename']\n        spec = None\n\n        if self.spectrograms and samplename in self.spectrograms:\n            spec = self.spectrograms[samplename]\n        elif not self.cfg.LOAD_DATA:\n            spec = process_audio_file(row['filepath'], self.cfg)\n\n        if spec is None:\n            spec = np.zeros(self.cfg.TARGET_SHAPE, dtype=np.float32)\n            if self.mode == \"train\":  # Only print warning during training\n                print(f\"Warning: Spectrogram for {samplename} not found and could not be generated\")\n\n        spec = torch.tensor(spec, dtype=torch.float32).unsqueeze(0)  # Add channel dimension\n\n        if self.mode == \"train\" and random.random() < self.cfg.aug_prob:\n            spec = self.apply_spec_augmentations(spec)\n        \n        target = self.encode_label(row['primary_label'])\n        \n        if 'secondary_labels' in row and row['secondary_labels'] not in [[''], None, np.nan]:\n            if isinstance(row['secondary_labels'], str):\n                secondary_labels = eval(row['secondary_labels'])\n            else:\n                secondary_labels = row['secondary_labels']\n            \n            for label in secondary_labels:\n                if label in self.label_to_idx:\n                    target[self.label_to_idx[label]] = 1.0\n        \n        return {\n            'melspec': spec, \n            'target': torch.tensor(target, dtype=torch.float32),\n            'filename': row['filename']\n        }\n   #-------------------------------------------------------------------------------------------------------- \n    def apply_spec_augmentations(self, spec):\n        \"\"\"Apply augmentations to spectrogram\"\"\"\n    \n        # Time masking (horizontal stripes)\n        if random.random() < 0.5:\n            num_masks = random.randint(1, 3)\n            for _ in range(num_masks):\n                width = random.randint(5, 20)\n                start = random.randint(0, spec.shape[2] - width)\n                spec[0, :, start:start+width] = 0\n        \n        # Frequency masking (vertical stripes)\n        if random.random() < 0.5:\n            num_masks = random.randint(1, 3)\n            for _ in range(num_masks):\n                height = random.randint(5, 20)\n                start = random.randint(0, spec.shape[1] - height)\n                spec[0, start:start+height, :] = 0\n        \n        # Random brightness/contrast\n        if random.random() < 0.5:\n            gain = random.uniform(0.8, 1.2)\n            bias = random.uniform(-0.1, 0.1)\n            spec = spec * gain + bias\n            spec = torch.clamp(spec, 0, 1) \n            \n        return spec\n    \n    def encode_label(self, label):\n        \"\"\"Encode label to one-hot vector\"\"\"\n        target = np.zeros(self.num_classes)\n        if label in self.label_to_idx:\n            target[self.label_to_idx[label]] = 1.0\n        return target","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-05-14T10:52:30.016433Z","iopub.execute_input":"2025-05-14T10:52:30.016621Z","iopub.status.idle":"2025-05-14T10:52:30.028035Z","shell.execute_reply.started":"2025-05-14T10:52:30.016605Z","shell.execute_reply":"2025-05-14T10:52:30.027375Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def collate_fn(batch):\n    \"\"\"Custom collate function to handle different sized spectrograms\"\"\"\n    batch = [item for item in batch if item is not None]\n    if len(batch) == 0:\n        return {}\n        \n    result = {key: [] for key in batch[0].keys()}\n    \n    for item in batch:\n        for key, value in item.items():\n            result[key].append(value)\n    \n    for key in result:\n        if key == 'target' and isinstance(result[key][0], torch.Tensor):\n            result[key] = torch.stack(result[key])\n        elif key == 'melspec' and isinstance(result[key][0], torch.Tensor):\n            shapes = [t.shape for t in result[key]]\n            if len(set(str(s) for s in shapes)) == 1:\n                result[key] = torch.stack(result[key])\n    \n    return result","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-05-14T10:52:33.439266Z","iopub.execute_input":"2025-05-14T10:52:33.439550Z","iopub.status.idle":"2025-05-14T10:52:33.445402Z","shell.execute_reply.started":"2025-05-14T10:52:33.439528Z","shell.execute_reply":"2025-05-14T10:52:33.444440Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Model Definition","metadata":{}},{"cell_type":"code","source":"class BirdCLEFModel(nn.Module):\n    def __init__(self, cfg):\n        super().__init__()\n        self.cfg = cfg\n\n        # 从分类法CSV文件中读取数据，获取类别数量并设置到配置中\n        taxonomy_df = pd.read_csv(cfg.taxonomy_csv)\n        cfg.num_classes = len(taxonomy_df)\n\n\n        # 创建模型骨干网络，使用timm库创建指定名称的模型\n        # pretrained表示是否加载预训练权重\n        # in_chans表示输入通道数\n        # drop_rate和drop_path_rate是模型的一些超参数\n        self.backbone = timm.create_model(\n            cfg.model_name,\n            pretrained=cfg.pretrained,\n            in_chans=cfg.in_channels,\n            drop_rate=0.2,\n            drop_path_rate=0.2\n        )\n\n        # 根据模型名称进行不同的处理，获取骨干网络输出特征维度并修改骨干网络的分类层\n        if 'efficientnet' in cfg.model_name:\n            backbone_out = self.backbone.classifier.in_features\n            self.backbone.classifier = nn.Identity()\n        elif 'resnet' in cfg.model_name:\n            backbone_out = self.backbone.fc.in_features\n            self.backbone.fc = nn.Identity()\n        else:\n            backbone_out = self.backbone.get_classifier().in_features\n            self.backbone.reset_classifier(0, '')\n\n        # 创建自适应平均池化层，将输入特征图池化到固定大小（这里是1x1）\n        self.pooling = nn.AdaptiveAvgPool2d(1)\n\n        # 保存骨干网络输出的特征维度\n        self.feat_dim = backbone_out\n\n        # 创建最终的分类层，将骨干网络输出特征映射到类别数量对应的维度\n        self.classifier = nn.Linear(backbone_out, cfg.num_classes)\n\n        # 检查配置中是否启用了mixup数据增强（根据mixup_alpha是否大于0判断）\n        self.mixup_enabled = hasattr(cfg, 'mixup_alpha') and cfg.mixup_alpha > 0\n        if self.mixup_enabled:\n            self.mixup_alpha = cfg.mixup_alpha\n            \n    def forward(self, x, targets=None):\n\n        # 如果处于训练模式且启用了mixup数据增强，并且有目标标签\n        if self.training and self.mixup_enabled and targets is not None:\n            # 执行mixup操作，混合输入数据和目标标签\n            mixed_x, targets_a, targets_b, lam = self.mixup_data(x, targets)\n            x = mixed_x\n        else:\n            targets_a, targets_b, lam = None, None, None\n\n        # 将输入数据传入骨干网络，获取骨干网络的输出特征\n        features = self.backbone(x)\n\n        \n        if isinstance(features, dict):\n            features = features['features']\n\n        # 如果特征张量是4维的（图像数据常见的维度形式），进行池化操作并展平特征\n        if len(features.shape) == 4:\n            features = self.pooling(features)\n            features = features.view(features.size(0), -1)\n        \n        logits = self.classifier(features)\n\n        # 如果处于训练模式且启用了mixup数据增强，并且有目标标签\n        if self.training and self.mixup_enabled and targets is not None:\n            # 使用mixup后的目标标签计算损失\n            loss = self.mixup_criterion(F.binary_cross_entropy_with_logits, \n                                       logits, targets_a, targets_b, lam)\n            return logits, loss\n            \n        return logits\n    \n    def mixup_data(self, x, targets):\n        \"\"\"Applies mixup to the data batch\"\"\"\n        batch_size = x.size(0)\n\n        lam = np.random.beta(self.mixup_alpha, self.mixup_alpha)\n\n        indices = torch.randperm(batch_size).to(x.device)\n\n        mixed_x = lam * x + (1 - lam) * x[indices]\n        \n        return mixed_x, targets, targets[indices], lam\n    \n    def mixup_criterion(self, criterion, pred, y_a, y_b, lam):\n        \"\"\"Applies mixup to the loss function\"\"\"\n        return lam * criterion(pred, y_a) + (1 - lam) * criterion(pred, y_b)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-05-14T10:52:33.816814Z","iopub.execute_input":"2025-05-14T10:52:33.817050Z","iopub.status.idle":"2025-05-14T10:52:33.826656Z","shell.execute_reply.started":"2025-05-14T10:52:33.817030Z","shell.execute_reply":"2025-05-14T10:52:33.825898Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Training Utilities\nWe are configuring our optimization strategy with the AdamW optimizer, cosine scheduling, and the BCEWithLogitsLoss criterion.","metadata":{}},{"cell_type":"code","source":"def get_optimizer(model, cfg):\n  \n    if cfg.optimizer == 'Adam':\n        optimizer = optim.Adam(\n            model.parameters(),\n            lr=cfg.lr,\n            weight_decay=cfg.weight_decay\n        )\n    elif cfg.optimizer == 'AdamW':\n        optimizer = optim.AdamW(\n            model.parameters(),\n            lr=cfg.lr,\n            weight_decay=cfg.weight_decay\n        )\n    elif cfg.optimizer == 'SGD':\n        optimizer = optim.SGD(\n            model.parameters(),\n            lr=cfg.lr,\n            momentum=0.9,\n            weight_decay=cfg.weight_decay\n        )\n    else:\n        raise NotImplementedError(f\"Optimizer {cfg.optimizer} not implemented\")\n        \n    return optimizer\n\ndef get_scheduler(optimizer, cfg):\n   \n    if cfg.scheduler == 'CosineAnnealingLR':\n        scheduler = lr_scheduler.CosineAnnealingLR(\n            optimizer,\n            T_max=cfg.T_max,\n            eta_min=cfg.min_lr\n        )\n    elif cfg.scheduler == 'ReduceLROnPlateau':\n        scheduler = lr_scheduler.ReduceLROnPlateau(\n            optimizer,\n            mode='min',\n            factor=0.5,\n            patience=2,\n            min_lr=cfg.min_lr,\n            verbose=True\n        )\n    elif cfg.scheduler == 'StepLR':\n        scheduler = lr_scheduler.StepLR(\n            optimizer,\n            step_size=cfg.epochs // 3,\n            gamma=0.5\n        )\n    elif cfg.scheduler == 'OneCycleLR':\n        scheduler = None  \n    else:\n        scheduler = None\n        \n    return scheduler\n\ndef get_criterion(cfg):\n \n    if cfg.criterion == 'BCEWithLogitsLoss':\n        criterion = nn.BCEWithLogitsLoss()\n    else:\n        raise NotImplementedError(f\"Criterion {cfg.criterion} not implemented\")\n        \n    return criterion","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-05-14T10:52:36.051849Z","iopub.execute_input":"2025-05-14T10:52:36.052143Z","iopub.status.idle":"2025-05-14T10:52:36.058667Z","shell.execute_reply.started":"2025-05-14T10:52:36.052101Z","shell.execute_reply":"2025-05-14T10:52:36.057831Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Training Loop","metadata":{}},{"cell_type":"code","source":"def train_one_epoch(model, loader, optimizer, criterion, device, scheduler=None):\n    \n    model.train()\n    losses = []\n    all_targets = []\n    all_outputs = []\n    \n    pbar = tqdm(enumerate(loader), total=len(loader), desc=\"Training\")\n    \n    for step, batch in pbar:\n    \n        if isinstance(batch['melspec'], list):\n            batch_outputs = []\n            batch_losses = []\n            \n            for i in range(len(batch['melspec'])):\n                inputs = batch['melspec'][i].unsqueeze(0).to(device)\n                target = batch['target'][i].unsqueeze(0).to(device)\n                \n                optimizer.zero_grad()\n                output = model(inputs)\n                loss = criterion(output, target)\n                loss.backward()\n                \n                batch_outputs.append(output.detach().cpu())\n                batch_losses.append(loss.item())\n            \n            optimizer.step()\n            outputs = torch.cat(batch_outputs, dim=0).numpy()\n            loss = np.mean(batch_losses)\n            targets = batch['target'].numpy()\n            \n        else:\n            inputs = batch['melspec'].to(device)\n            targets = batch['target'].to(device)\n            \n            optimizer.zero_grad()\n            outputs = model(inputs)\n            \n            if isinstance(outputs, tuple):\n                outputs, loss = outputs  \n            else:\n                loss = criterion(outputs, targets)\n                \n            loss.backward()\n            optimizer.step()\n            \n            outputs = outputs.detach().cpu().numpy()\n            targets = targets.detach().cpu().numpy()\n        \n        if scheduler is not None and isinstance(scheduler, lr_scheduler.OneCycleLR):\n            scheduler.step()\n            \n        all_outputs.append(outputs)\n        all_targets.append(targets)\n        losses.append(loss if isinstance(loss, float) else loss.item())\n        \n        pbar.set_postfix({\n            'train_loss': np.mean(losses[-10:]) if losses else 0,\n            'lr': optimizer.param_groups[0]['lr']\n        })\n    \n    all_outputs = np.concatenate(all_outputs)\n    all_targets = np.concatenate(all_targets)\n    auc = calculate_auc(all_targets, all_outputs)\n    avg_loss = np.mean(losses)\n    \n    return avg_loss, auc\n\ndef validate(model, loader, criterion, device):\n   \n    model.eval()\n    losses = []\n    all_targets = []\n    all_outputs = []\n    \n    with torch.no_grad():\n        for batch in tqdm(loader, desc=\"Validation\"):\n            if isinstance(batch['melspec'], list):\n                batch_outputs = []\n                batch_losses = []\n                \n                for i in range(len(batch['melspec'])):\n                    inputs = batch['melspec'][i].unsqueeze(0).to(device)\n                    target = batch['target'][i].unsqueeze(0).to(device)\n                    \n                    output = model(inputs)\n                    loss = criterion(output, target)\n                    \n                    batch_outputs.append(output.detach().cpu())\n                    batch_losses.append(loss.item())\n                \n                outputs = torch.cat(batch_outputs, dim=0).numpy()\n                loss = np.mean(batch_losses)\n                targets = batch['target'].numpy()\n                \n            else:\n                inputs = batch['melspec'].to(device)\n                targets = batch['target'].to(device)\n                \n                outputs = model(inputs)\n                loss = criterion(outputs, targets)\n                \n                outputs = outputs.detach().cpu().numpy()\n                targets = targets.detach().cpu().numpy()\n            \n            all_outputs.append(outputs)\n            all_targets.append(targets)\n            losses.append(loss if isinstance(loss, float) else loss.item())\n    \n    all_outputs = np.concatenate(all_outputs)\n    all_targets = np.concatenate(all_targets)\n    \n    auc = calculate_auc(all_targets, all_outputs)\n    avg_loss = np.mean(losses)\n    \n    return avg_loss, auc\n\ndef calculate_auc(targets, outputs):\n  \n    num_classes = targets.shape[1]\n    aucs = []\n    \n    probs = 1 / (1 + np.exp(-outputs))\n    \n    for i in range(num_classes):\n        \n        if np.sum(targets[:, i]) > 0:\n            class_auc = roc_auc_score(targets[:, i], probs[:, i])\n            aucs.append(class_auc)\n    \n    return np.mean(aucs) if aucs else 0.0","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-05-14T10:52:36.397207Z","iopub.execute_input":"2025-05-14T10:52:36.397433Z","iopub.status.idle":"2025-05-14T10:52:36.410333Z","shell.execute_reply.started":"2025-05-14T10:52:36.397414Z","shell.execute_reply":"2025-05-14T10:52:36.409427Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Training!","metadata":{}},{"cell_type":"code","source":"def run_training(df, cfg):\n    \"\"\"Training function that can either use pre-computed spectrograms or generate them on-the-fly\"\"\"\n\n    taxonomy_df = pd.read_csv(cfg.taxonomy_csv)\n    species_ids = taxonomy_df['primary_label'].tolist()\n    cfg.num_classes = len(species_ids)\n    \n    if cfg.debug:\n        cfg.update_debug_settings()\n\n    spectrograms = None\n    if cfg.LOAD_DATA:\n        print(\"Loading pre-computed mel spectrograms from NPY file...\")\n        try:\n            spectrograms = np.load(cfg.spectrogram_npy, allow_pickle=True).item()\n            print(f\"Loaded {len(spectrograms)} pre-computed mel spectrograms\")\n        except Exception as e:\n            print(f\"Error loading pre-computed spectrograms: {e}\")\n            print(\"Will generate spectrograms on-the-fly instead.\")\n            cfg.LOAD_DATA = False\n    \n    if not cfg.LOAD_DATA:\n        print(\"Will generate spectrograms on-the-fly during training.\")\n        if 'filepath' not in df.columns:\n            df['filepath'] = cfg.train_datadir + '/' + df.filename\n        if 'samplename' not in df.columns:\n            df['samplename'] = df.filename.map(lambda x: x.split('/')[0] + '-' + x.split('/')[-1].split('.')[0])\n        \n    skf = StratifiedKFold(n_splits=7, shuffle=True, random_state=cfg.seed)\n    \n    best_scores = []\n\n    \n    for fold, (train_idx, val_idx) in enumerate(skf.split(df, df['primary_label'])):\n        if fold not in cfg.selected_folds:\n            continue\n            \n        print(f'\\n{\"=\"*30} Fold {fold} {\"=\"*30}')\n        \n        train_df = df.iloc[train_idx].reset_index(drop=True)\n        val_df = df.iloc[val_idx].reset_index(drop=True)\n        \n        print(f'Training set: {len(train_df)} samples')\n        print(f'Validation set: {len(val_df)} samples')\n        \n        train_dataset = BirdCLEFDatasetFromNPY(train_df, cfg, spectrograms=spectrograms, mode='train')\n        val_dataset = BirdCLEFDatasetFromNPY(val_df, cfg, spectrograms=spectrograms, mode='valid')\n        \n        train_loader = DataLoader(\n            train_dataset, \n            batch_size=cfg.batch_size, \n            shuffle=True, \n            num_workers=cfg.num_workers,\n            pin_memory=True,\n            collate_fn=collate_fn,\n            drop_last=True\n        )\n        \n        val_loader = DataLoader(\n            val_dataset, \n            batch_size=cfg.batch_size, \n            shuffle=False, \n            num_workers=cfg.num_workers,\n            pin_memory=True,\n            collate_fn=collate_fn\n        )\n        \n        model = BirdCLEFModel(cfg).to(cfg.device)\n        optimizer = get_optimizer(model, cfg)\n        criterion = get_criterion(cfg)\n        \n        if cfg.scheduler == 'OneCycleLR':\n            scheduler = lr_scheduler.OneCycleLR(\n                optimizer,\n                max_lr=cfg.lr,\n                steps_per_epoch=len(train_loader),\n                epochs=cfg.epochs,\n                pct_start=0.1\n            )\n        else:\n            scheduler = get_scheduler(optimizer, cfg)\n        \n        best_auc = 0\n        best_epoch = 0\n\n        \n        for epoch in range(cfg.epochs):\n            print(f\"\\nEpoch {epoch+1}/{cfg.epochs}\")\n            \n            train_loss, train_auc = train_one_epoch(\n                model, \n                train_loader, \n                optimizer, \n                criterion, \n                cfg.device,\n                scheduler if isinstance(scheduler, lr_scheduler.OneCycleLR) else None\n            )\n            \n            val_loss, val_auc = validate(model, val_loader, criterion, cfg.device)\n\n            if scheduler is not None and not isinstance(scheduler, lr_scheduler.OneCycleLR):\n                if isinstance(scheduler, lr_scheduler.ReduceLROnPlateau):\n                    scheduler.step(val_loss)\n                else:\n                    scheduler.step()\n\n            print(f\"Train Loss: {train_loss:.4f}, Train AUC: {train_auc:.4f}\")\n            print(f\"Val Loss: {val_loss:.4f}, Val AUC: {val_auc:.4f}\")\n\n            if val_auc > best_auc:\n                best_auc = val_auc\n                best_epoch = epoch + 1\n                print(f\"New best AUC: {best_auc:.4f} at epoch {best_epoch}\")\n\n                torch.save({\n                    'model_state_dict': model.state_dict(),\n                    'optimizer_state_dict': optimizer.state_dict(),\n                    'scheduler_state_dict': scheduler.state_dict() if scheduler else None,\n                    'epoch': epoch,\n                    'val_auc': val_auc,\n                    'train_auc': train_auc,\n                    'cfg': cfg\n                }, f\"model_fold{fold}.pth\")\n\n        \n        best_scores.append(best_auc)\n        print(f\"\\nBest AUC for fold {fold}: {best_auc:.4f} at epoch {best_epoch}\")\n        \n        # Clear memory\n        del model, optimizer, scheduler, train_loader, val_loader\n        torch.cuda.empty_cache()\n        gc.collect()\n    \n    print(\"\\n\" + \"=\"*60)\n    print(\"Cross-Validation Results:\")\n    for fold, score in enumerate(best_scores):\n        print(f\"Fold {cfg.selected_folds[fold]}: {score:.4f}\")\n    print(f\"Mean AUC: {np.mean(best_scores):.4f}\")\n    print(\"=\"*60)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-05-14T10:52:39.956255Z","iopub.execute_input":"2025-05-14T10:52:39.956548Z","iopub.status.idle":"2025-05-14T10:52:39.969044Z","shell.execute_reply.started":"2025-05-14T10:52:39.956526Z","shell.execute_reply":"2025-05-14T10:52:39.968303Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"if __name__ == \"__main__\":\n    import time\n    \n    print(\"\\nLoading training data...\")\n    train_df = pd.read_csv(cfg.train_csv)\n    taxonomy_df = pd.read_csv(cfg.taxonomy_csv)\n\n    print(\"\\nStarting training...\")\n    print(f\"LOAD_DATA is set to {cfg.LOAD_DATA}\")\n    if cfg.LOAD_DATA:\n        print(\"Using pre-computed mel spectrograms from NPY file\")\n    else:\n        print(\"Will generate spectrograms on-the-fly during training\")\n    \n    run_training(train_df, cfg)\n    \n    print(\"\\nTraining complete!\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-05-14T10:52:40.190298Z","iopub.execute_input":"2025-05-14T10:52:40.190504Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null}]}