{"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,"isSourceIdPinned":false,"sourceType":"competition"},{"sourceId":11060723,"sourceType":"datasetVersion","datasetId":6891568},{"sourceId":11269980,"sourceType":"datasetVersion","datasetId":7044820},{"sourceId":11471161,"sourceType":"datasetVersion","datasetId":7188855}],"dockerImageVersionId":30918,"isInternetEnabled":false,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"# **BirdCLEF 2025 Inference Notebook**\nThis notebook runs inference on BirdCLEF 2025 test soundscapes and generates a submission file. It supports both single model inference and ensemble inference with multiple models. You can find the pre-processing and training processes in the following notebooks:\n\n- [Transforming Audio-to-Mel Spec. | BirdCLEF'25](https://www.kaggle.com/code/kadircandrisolu/transforming-audio-to-mel-spec-birdclef-25)  \n- [EfficientNet B0 Pytorch [Train] | BirdCLEF'25](https://www.kaggle.com/code/kadircandrisolu/efficientnet-b0-pytorch-train-birdclef-25)\n\n**Features**\n- Audio Preprocessing\n- Test-Time Augmentation (TTA)","metadata":{}},{"cell_type":"code","source":"import os\nimport gc\nimport warnings\nimport logging\nimport time\nimport math\nimport cv2\nfrom pathlib import Path\n\nimport numpy as np\nimport pandas as pd\nimport librosa\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport timm\nfrom tqdm.auto import tqdm\n\nwarnings.filterwarnings(\"ignore\")\nlogging.basicConfig(level=logging.ERROR)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-04-22T14:23:13.736415Z","iopub.execute_input":"2025-04-22T14:23:13.736643Z","iopub.status.idle":"2025-04-22T14:23:26.497870Z","shell.execute_reply.started":"2025-04-22T14:23:13.736611Z","shell.execute_reply":"2025-04-22T14:23:26.496953Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class CFG:\n \n    test_soundscapes = '/kaggle/input/birdclef-2025/test_soundscapes'\n    submission_csv = '/kaggle/input/birdclef-2025/sample_submission.csv'\n    taxonomy_csv = '/kaggle/input/birdclef-2025/taxonomy.csv'\n    model_path = '/kaggle/input/0419try'  #注意这里写自己上传的dataset的目录，而不是模型文件本身（‘/kaggle/input/baseline/model_fold0.pth’就会报错）\n    \n    # Audio parameters\n    FS = 32000  \n    WINDOW_SIZE = 5  \n    \n    # Mel spectrogram parameters\n    N_FFT = 1024\n    HOP_LENGTH = 512\n    N_MELS = 128\n    FMIN = 50\n    FMAX = 14000\n    TARGET_SHAPE = (256, 256)\n    num_classes = 264\n    backbone = 'tf_efficientnetv2_s_in21k'\n    mixup_version = 'v1'\n    train_period = 5\n    val_period = 5   \n\n\n\n    \n    model_name = 'efficientnet_b0'\n    in_channels = 1\n    device = 'cpu'  \n    \n    # Inference parameters\n    batch_size = 16\n    use_tta = False  \n    tta_count = 3   \n    threshold = 0.5\n    \n    use_specific_folds = False  # If False, use all found models\n    folds = [0, 1]  # Used only if use_specific_folds is True\n    \n    debug = False\n    debug_count = 3\n\ncfg = CFG()\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-04-22T14:23:26.498862Z","iopub.execute_input":"2025-04-22T14:23:26.499185Z","iopub.status.idle":"2025-04-22T14:23:26.504287Z","shell.execute_reply.started":"2025-04-22T14:23:26.499151Z","shell.execute_reply":"2025-04-22T14:23:26.503305Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"#model1\n# -*- coding: utf-8 -*-\n\"\"\"\nref from BirdCLEF2023 rank4 solution\n\"\"\"\nimport torch\nimport torch.nn as nn\nimport numpy as np\nimport timm\nimport torch.nn.functional as F\n\n#from models.aug import MixupV2, Mixup\n\n# ⚠️ mock 一个假的 Mixup，避免 inference 阶段出错\nclass Mixup:\n    def __init__(self, *args, **kwargs):\n        pass\n    def __call__(self, x, y):\n        return x, y, torch.ones_like(y), y\nclass MixupV2(Mixup):\n    pass\n# ✅ 在 AttModel 定义前添加：\nclass GeM(nn.Module):\n    def __init__(self, p=3, eps=1e-6):\n        super().__init__()\n        self.p = nn.Parameter(torch.ones(1) * p)\n        self.eps = eps\n\n    def forward(self, x):\n        return F.adaptive_avg_pool2d(x.clamp(min=self.eps).pow(self.p), (1, 1)).pow(1. / self.p)\n\n\nclass AttModel(nn.Module):\n    def __init__(self,\n                 backbone='tf_efficientnetv2_s_in21k',\n                 mixup_version='v1',\n                 load_pretrained_from=None,\n                 num_classes=CFG.num_classes):\n        super().__init__()\n\n        # backbone = CFG.backbone\n        train_period = CFG.train_period = 5\n        infer_period = CFG.val_period\n\n        base_model = timm.create_model(\n            backbone,\n            features_only=False,\n            pretrained=False,  # ✅ timm 会自动下载 safetensors 权重\n            in_chans=3,\n            )\n\n\n        n = 2\n        if 'tiny_vit' in backbone:\n            n = 1\n        if 'efficientvit' in backbone:\n            n = 1\n        if 'caformer' in backbone:\n            n = 1\n        layers = list(base_model.children())[:-n]\n        self.backbone = nn.Sequential(*layers)\n        if \"efficientnet\" in backbone:\n            dense_input = base_model.num_features\n        elif hasattr(base_model, \"fc\"):\n            dense_input = base_model.fc.in_features\n        else:\n            dense_input = base_model.feature_info[-1][\"num_chs\"]\n        self.dense_input = dense_input\n        self.train_period = train_period\n        self.infer_period = infer_period\n        self.factor = int(self.train_period / self.infer_period)\n        self.mixup_version = mixup_version\n        if mixup_version == 'v2' or mixup_version == 'v3':\n            self.mixup = MixupV2(mix_range=(0.3, 0.7), add_label=True)\n        else:\n            self.mixup = Mixup(mix_beta=1)\n        self.global_pool = GeM()\n        self.dropouts = nn.ModuleList([nn.Dropout(p) for p in np.linspace(0.1, 0.5, 5)])\n        self.head = nn.Linear(1280, 206)\n        \n        \n\n\n    def forward(self, input, return_all=True, with_mix_up=True):\n\n        image = input[\"image\"]\n        y = input[\"loss_target\"]\n        teacher_preds = input[\"teacher_preds\"]\n        weight = torch.ones_like(y)  # ✅ 防止没有进入 mixup 就 return 时出错\n        \n        if y.ndim == 1:\n            y = y.view(-1, CFG.num_classes)\n\n        '''\n        if self.training and with_mix_up:\n            if np.random.random() <= 0.5:\n                if self.mixup_version == 'v3':\n                    n_mix = np.random.randint(1, 3)\n                    for _ in range(n_mix):\n                        image, y, weight, teacher_preds = self.mixup(image, y)\n                else:\n                    # ✅ 仅在 image 和 y 数量对得上的时候才 mixup\n                    if image.size(0) == y.size(0):\n                        image, y, weight, teacher_preds = self.mixup(image, y)\n                    else:\n                        # ⚠️ 跳过 mixup，保持一致性\n                        weight = torch.ones_like(y)\n                        teacher_preds = y.clone()\n'''\n\n\n        x = self.backbone(image)\n\n        x = self.global_pool(x)\n        x = x[:, :, 0, 0]\n        logit = sum([self.head(dropout(x)) for dropout in self.dropouts]) / 5\n\n        if return_all:\n            return {\"logit\": logit, \"target\": y, \"rating\": weight, \"teacher_preds\": teacher_preds,\n                    \"image\": image}\n\n        return {\"logit\": logit, }\n\n\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-04-22T14:23:26.505166Z","iopub.execute_input":"2025-04-22T14:23:26.505525Z","iopub.status.idle":"2025-04-22T14:23:26.565704Z","shell.execute_reply.started":"2025-04-22T14:23:26.505493Z","shell.execute_reply":"2025-04-22T14:23:26.564853Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"print(f\"Using device: {cfg.device}\")\nprint(f\"Loading taxonomy data...\")\ntaxonomy_df = pd.read_csv(cfg.taxonomy_csv)\nspecies_ids = taxonomy_df['primary_label'].tolist()\nnum_classes = len(species_ids)\nprint(f\"Number of classes: {num_classes}\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-04-22T14:23:26.567864Z","iopub.execute_input":"2025-04-22T14:23:26.568093Z","iopub.status.idle":"2025-04-22T14:23:26.603191Z","shell.execute_reply.started":"2025-04-22T14:23:26.568074Z","shell.execute_reply":"2025-04-22T14:23:26.602540Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class BirdCLEFModel(nn.Module):\n    def __init__(self, cfg, num_classes):\n        super().__init__()\n        self.cfg = cfg\n        \n        self.backbone = timm.create_model(\n            cfg.model_name,\n            pretrained=False,  \n            in_chans=cfg.in_channels,\n            drop_rate=0.0,    \n            drop_path_rate=0.0\n        )\n        \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        self.pooling = nn.AdaptiveAvgPool2d(1)\n        self.feat_dim = backbone_out\n        self.classifier = nn.Linear(backbone_out, num_classes)\n        \n    def forward(self, x):\n        features = self.backbone(x)\n        \n        if isinstance(features, dict):\n            features = features['features']\n            \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        return logits\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-04-22T14:23:26.604409Z","iopub.execute_input":"2025-04-22T14:23:26.604606Z","iopub.status.idle":"2025-04-22T14:23:26.610688Z","shell.execute_reply.started":"2025-04-22T14:23:26.604589Z","shell.execute_reply":"2025-04-22T14:23:26.609889Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def audio2melspec(audio_data, cfg):\n    \"\"\"Convert audio data to mel spectrogram\"\"\"\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    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    mel_spec_db = librosa.power_to_db(mel_spec, ref=np.max)\n    mel_spec_norm = (mel_spec_db - mel_spec_db.min()) / (mel_spec_db.max() - mel_spec_db.min() + 1e-8)\n    \n    # 👉 Convert to 3-channel mel spectrogram\n    mel_spec = np.transpose(mel_spec, (1, 2, 0))  # (F, T, 3)\n    mel_spec = cv2.resize(mel_spec, cfg.TARGET_SHAPE, interpolation=cv2.INTER_LINEAR)\n    mel_spec = np.transpose(mel_spec, (2, 0, 1))  # 回到 (3, H, W)\n\n    return mel_spec_norm\n\ndef process_audio_segment(audio_data, cfg):\n    \"\"\"Process audio segment to get mel spectrogram\"\"\"\n    if len(audio_data) < cfg.FS * cfg.WINDOW_SIZE:\n        audio_data = np.pad(audio_data, \n                          (0, cfg.FS * cfg.WINDOW_SIZE - len(audio_data)), \n                          mode='constant')\n    \n    mel_spec = audio2melspec(audio_data, cfg)  \n\n\n    # Resize if needed\n    if mel_spec.shape != Cfg.TARGET_SHAPE:\n        mel_spec = cv2.resize(mel_spec, cfg.TARGET_SHAPE, interpolation=cv2.INTER_LINEAR)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-04-22T14:23:26.611384Z","iopub.execute_input":"2025-04-22T14:23:26.611562Z","iopub.status.idle":"2025-04-22T14:23:26.628494Z","shell.execute_reply.started":"2025-04-22T14:23:26.611547Z","shell.execute_reply":"2025-04-22T14:23:26.627764Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def find_model_files(cfg):\n    \"\"\"\n    Find all .pth model files in the specified model directory\n    \"\"\"\n    model_files = []\n    \n    model_dir = Path(cfg.model_path)\n    \n    for path in model_dir.glob('**/*.pth'):\n        model_files.append(str(path))\n    \n    return model_files\n\ndef load_models(cfg, num_classes):\n    \"\"\"\n    Load all found model files and prepare them for ensemble\n    \"\"\"\n    models = []\n    \n    model_files = find_model_files(cfg)\n    \n    if not model_files:\n        print(f\"Warning: No model files found under {cfg.model_path}!\")\n        return models\n    \n    print(f\"Found a total of {len(model_files)} model files.\")\n\n    \n    if cfg.use_specific_folds:\n        filtered_files = []\n        for fold in cfg.folds:\n            fold_files = [f for f in model_files if f\"fold{fold}\" in f]\n            filtered_files.extend(fold_files)\n        model_files = filtered_files\n        print(f\"Using {len(model_files)} model files for the specified folds ({cfg.folds}).\")\n    \n\n\n    for model_path in model_files:\n        try:\n            #print(f\"🔄 Loading model: {model_path}\")\n\n        # ✅ ① 创建 AttModel（用训练时一致的参数）\n            model = AttModel(\n                backbone=cfg.backbone,\n                mixup_version=cfg.mixup_version,\n                load_pretrained_from=None\n            )\n    \n        # ✅ ② 加载权重（直接是 state_dict，不是 checkpoint 包）\n            state_dict = torch.load(model_path, map_location=torch.device(cfg.device))\n            model.load_state_dict(state_dict)\n\n        # ✅ ③ 模型迁移到设备并设置 eval 模式\n            model = model.to(cfg.device)\n            model.eval()\n\n            models.append(model)\n            print(f\"✅ Loaded model from {model_path}\")\n\n        except Exception as e:\n            print(f\"❌ Error loading model {model_path}: {e}\")\n            pass\n\n    print(f\"✅ Returning {len(models)} loaded models\")\n    return models\n\n\ndef predict_on_spectrogram(audio_path, models, cfg, species_ids):\n    \"\"\"Process a single audio file and predict species presence for each 5-second segment\"\"\"\n    predictions = []\n    row_ids = []\n    soundscape_id = Path(audio_path).stem\n    \n    try:\n        print(f\"Processing {soundscape_id}\")\n        audio_data, _ = librosa.load(audio_path, sr=cfg.FS)\n        \n        total_segments = int(len(audio_data) / (cfg.FS * cfg.WINDOW_SIZE))\n        \n        for segment_idx in range(total_segments):\n            start_sample = segment_idx * cfg.FS * cfg.WINDOW_SIZE\n            end_sample = start_sample + cfg.FS * cfg.WINDOW_SIZE\n            segment_audio = audio_data[start_sample:end_sample]\n            \n            end_time_sec = (segment_idx + 1) * cfg.WINDOW_SIZE\n            row_id = f\"{soundscape_id}_{end_time_sec}\"\n            row_ids.append(row_id)\n\n            if cfg.use_tta:\n                all_preds = []\n                \n                for tta_idx in range(cfg.tta_count):\n                    mel_spec = process_audio_segment(segment_audio, cfg)\n                    mel_spec = apply_tta(mel_spec, tta_idx)\n\n                    mel_spec = torch.tensor(mel_spec, dtype=torch.float32).unsqueeze(0).unsqueeze(0)\n                    mel_spec = mel_spec.to(cfg.device)\n\n                    if len(models) == 1:\n                        with torch.no_grad():\n                            input_dict = {\n                                \"image\": mel_spec,\n                                \"loss_target\": torch.zeros((1, cfg.num_classes)).to(cfg.device),\n                                \"teacher_preds\": torch.zeros((1, cfg.num_classes)).to(cfg.device),\n                            }\n                            outputs = models[0](mel_spec)  \n\n                            probs = torch.sigmoid(outputs).cpu().numpy().squeeze()\n                            all_preds.append(probs)\n                    else:\n                        segment_preds = []\n                        for model in models:\n                            with torch.no_grad():\n                                input_dict = {\n                                    \"image\": mel_spec,\n                                    \"loss_target\": torch.zeros((1, cfg.num_classes)).to(cfg.device),\n                                    \"teacher_preds\": torch.zeros((1, cfg.num_classes)).to(cfg.device),\n                                }\n                                outputs = models[0](mel_spec) \n                                probs = torch.sigmoid(outputs).cpu().numpy().squeeze()\n                                segment_preds.append(probs)\n                        \n                        avg_preds = np.mean(segment_preds, axis=0)\n                        all_preds.append(avg_preds)\n\n                final_preds = np.mean(all_preds, axis=0)\n            else:\n                mel_spec = process_audio_segment(segment_audio, cfg)\n                \n                mel_spec = torch.tensor(mel_spec, dtype=torch.float32).unsqueeze(0).unsqueeze(0)\n                mel_spec = mel_spec.to(cfg.device)\n                \n                if len(models) == 1:\n                    with torch.no_grad():\n                        outputs = models[0](mel_spec)\n                        final_preds = torch.sigmoid(outputs).cpu().numpy().squeeze()\n                else:\n                    segment_preds = []\n                    for model in models:\n                        with torch.no_grad():\n                            outputs = model(mel_spec)\n                            probs = torch.sigmoid(outputs).cpu().numpy().squeeze()\n                            segment_preds.append(probs)\n\n                    final_preds = np.mean(segment_preds, axis=0)\n                    \n            predictions.append(final_preds)\n            \n    except Exception as e:\n        print(f\"Error processing {audio_path}: {e}\")\n    \n    return row_ids, predictions","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-04-22T14:23:26.629273Z","iopub.execute_input":"2025-04-22T14:23:26.629553Z","iopub.status.idle":"2025-04-22T14:23:26.652274Z","shell.execute_reply.started":"2025-04-22T14:23:26.629509Z","shell.execute_reply":"2025-04-22T14:23:26.651462Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def apply_tta(spec, tta_idx):\n    \"\"\"Apply test-time augmentation\"\"\"\n    if tta_idx == 0:\n        # Original spectrogram\n        return spec\n    elif tta_idx == 1:\n        # Time shift (horizontal flip)\n        return np.flip(spec, axis=1)\n    elif tta_idx == 2:\n        # Frequency shift (vertical flip)\n        return np.flip(spec, axis=0)\n    else:\n        return spec\n\ndef run_inference(cfg, models, species_ids):\n    \"\"\"Run inference on all test soundscapes\"\"\"\n    test_files = list(Path(cfg.test_soundscapes).glob('*.ogg'))\n    \n    if cfg.debug:\n        print(f\"Debug mode enabled, using only {cfg.debug_count} files\")\n        test_files = test_files[:cfg.debug_count]\n    \n    print(f\"Found {len(test_files)} test soundscapes\")\n\n    all_row_ids = []\n    all_predictions = []\n\n    for audio_path in tqdm(test_files):\n        row_ids, predictions = predict_on_spectrogram(str(audio_path), models, cfg, species_ids)\n        all_row_ids.extend(row_ids)\n        all_predictions.extend(predictions)\n    \n    return all_row_ids, all_predictions\n\ndef create_submission(row_ids, predictions, species_ids, cfg):\n    \"\"\"Create submission dataframe\"\"\"\n    print(\"Creating submission dataframe...\")\n\n    submission_dict = {'row_id': row_ids}\n    \n    for i, species in enumerate(species_ids):\n        submission_dict[species] = [pred[i] for pred in predictions]\n\n    submission_df = pd.DataFrame(submission_dict)\n\n    submission_df.set_index('row_id', inplace=True)\n\n    sample_sub = pd.read_csv(cfg.submission_csv, index_col='row_id')\n\n    missing_cols = set(sample_sub.columns) - set(submission_df.columns)\n    if missing_cols:\n        print(f\"Warning: Missing {len(missing_cols)} species columns in submission\")\n        for col in missing_cols:\n            submission_df[col] = 0.0\n\n    submission_df = submission_df[sample_sub.columns]\n\n    submission_df = submission_df.reset_index()\n    \n    return submission_df\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-04-22T14:23:26.653182Z","iopub.execute_input":"2025-04-22T14:23:26.653607Z","iopub.status.idle":"2025-04-22T14:23:26.676195Z","shell.execute_reply.started":"2025-04-22T14:23:26.653578Z","shell.execute_reply":"2025-04-22T14:23:26.675568Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def main():\n    start_time = time.time()\n    print(\"Starting BirdCLEF-2025 inference...\")\n    print(f\"TTA enabled: {cfg.use_tta} (variations: {cfg.tta_count if cfg.use_tta else 0})\")\n\n    models = load_models(cfg, num_classes)\n    \n    if not models:\n        print(\"No models found! Please check model paths.\")\n        return\n    \n    print(f\"Model usage: {'Single model' if len(models) == 1 else f'Ensemble of {len(models)} models'}\")\n\n    row_ids, predictions = run_inference(cfg, models, species_ids)\n\n    submission_df = create_submission(row_ids, predictions, species_ids, cfg)\n\n    submission_path = 'submission.csv'\n    submission_df.to_csv(submission_path, index=False)\n    print(f\"Submission saved to {submission_path}\")\n\n    end_time = time.time()\n    print(f\"Inference completed in {(end_time - start_time)/60:.2f} minutes\")\n    return submission_df  # ✅ return 出来！\n\n# ✅ Notebook 最后执行 main() 并显示结果\nsubmission_df = main()\nsubmission_df.head()  # ✅ 显示 DataFrame 供评测系统抓取\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-04-22T14:23:26.676884Z","iopub.execute_input":"2025-04-22T14:23:26.677068Z","iopub.status.idle":"2025-04-22T14:23:28.471425Z","shell.execute_reply.started":"2025-04-22T14:23:26.677053Z","shell.execute_reply":"2025-04-22T14:23:28.470728Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"if __name__ == \"__main__\":\n    main()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-04-22T14:23:28.472150Z","iopub.execute_input":"2025-04-22T14:23:28.472348Z","iopub.status.idle":"2025-04-22T14:23:29.343560Z","shell.execute_reply.started":"2025-04-22T14:23:28.472331Z","shell.execute_reply":"2025-04-22T14:23:29.342934Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\n\nprint(\"📁 test_soundscapes 路径是否存在：\", os.path.exists(CFG.test_soundscapes))\nprint(\"📂 里面的文件列表：\", os.listdir(CFG.test_soundscapes))\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-04-22T14:23:29.344464Z","iopub.execute_input":"2025-04-22T14:23:29.344740Z","iopub.status.idle":"2025-04-22T14:23:29.350193Z","shell.execute_reply.started":"2025-04-22T14:23:29.344692Z","shell.execute_reply":"2025-04-22T14:23:29.349319Z"}},"outputs":[],"execution_count":null}]}