{"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":"none","dataSources":[{"sourceId":91844,"databundleVersionId":11361821,"sourceType":"competition"},{"sourceId":12079451,"sourceType":"datasetVersion","datasetId":7534848,"isSourceIdPinned":true}],"dockerImageVersionId":30918,"isInternetEnabled":false,"language":"python","sourceType":"notebook","isGpuEnabled":false}},"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":"2025-06-06T11:20:12.696535Z","iopub.execute_input":"2025-06-06T11:20:12.696803Z","iopub.status.idle":"2025-06-06T11:20:24.702765Z","shell.execute_reply.started":"2025-06-06T11:20:12.696777Z","shell.execute_reply":"2025-06-06T11:20:24.701986Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def apply_power_to_low_ranked_cols(\n    p: np.ndarray,\n    top_k: int = 30,\n    exponent: float = 2.0,\n    inplace: bool = True\n) -> np.ndarray:\n    if not inplace:\n        p = p.copy()\n    tail_cols = np.argsort(-p.max(axis=0))[top_k:]\n    p[:, tail_cols] = p[:, tail_cols] ** exponent\n    return p","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-06T11:20:24.703707Z","iopub.execute_input":"2025-06-06T11:20:24.703960Z","iopub.status.idle":"2025-06-06T11:20:24.708505Z","shell.execute_reply.started":"2025-06-06T11:20:24.703939Z","shell.execute_reply":"2025-06-06T11:20:24.707820Z"}},"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/novoice-10sec-rms'  \n    \n    # Audio parameters\n    FS = 32000  \n    WINDOW_SIZE = 5\n    target_duration = 5  # length of desired inference segment\n    \n    # Mel spectrogram parameters\n    N_FFT = 1024\n    HOP_LENGTH = 256\n    N_MELS = 128\n    FMIN = 48\n    FMAX = 15000\n    TARGET_SHAPE = (256, 256)\n    \n    model_name = 'tf_efficientnetv2_b3.in21k'\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 = 0.7\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 = 5\n\ncfg = CFG()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-06T11:20:24.709507Z","iopub.execute_input":"2025-06-06T11:20:24.709867Z","iopub.status.idle":"2025-06-06T11:20:24.727166Z","shell.execute_reply.started":"2025-06-06T11:20:24.709824Z","shell.execute_reply":"2025-06-06T11:20:24.726312Z"}},"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":"2025-06-06T11:20:24.727970Z","iopub.execute_input":"2025-06-06T11:20:24.728228Z","iopub.status.idle":"2025-06-06T11:20:24.758903Z","shell.execute_reply.started":"2025-06-06T11:20:24.728206Z","shell.execute_reply":"2025-06-06T11:20:24.758029Z"}},"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=0.2,    \n            drop_path_rate=0.2\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":"2025-06-06T11:20:24.759814Z","iopub.execute_input":"2025-06-06T11:20:24.760179Z","iopub.status.idle":"2025-06-06T11:20:24.768914Z","shell.execute_reply.started":"2025-06-06T11:20:24.760147Z","shell.execute_reply":"2025-06-06T11:20:24.767987Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def load_sample(path, cfg):\n    import soundfile as sf\n    audio, orig_sr = sf.read(path, dtype=\"float32\")\n\n    seconds = []\n    audio_length = cfg.FS * cfg.target_duration\n    step = audio_length\n    for i in range(audio_length, len(audio) + step, step):\n        start = max(0, i - audio_length)\n        end = start + audio_length\n        if end > len(audio):\n            pass\n        else:\n            seconds.append(int(end / cfg.FS))\n\n    # pad beginning and end with zeros via concatenation\n    audio = np.concatenate([audio, audio, audio])\n    audios = []\n\n    for i, second in enumerate(seconds):\n        end_seconds = int(second)\n        start_seconds = int(end_seconds - cfg.target_duration)\n\n        end_index = int(cfg.FS * (end_seconds + (cfg.train_duration - cfg.target_duration) / 2)) + len(audio) // 3\n        start_index = int(cfg.FS * (start_seconds - (cfg.train_duration - cfg.target_duration) / 2)) + len(audio) // 3\n\n        y = audio[start_index:end_index].astype(np.float32)\n\n        # zero-pad at edges\n        start_pad = int(cfg.FS * (cfg.train_duration - cfg.target_duration) / 2)\n        end_pad = start_pad\n        if i == 0:\n            y[:start_pad] = 0\n        elif i == (len(seconds) - 1):\n            y[-end_pad:] = 0\n\n        audios.append(y)\n\n    return audios\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-06T11:20:24.771133Z","iopub.execute_input":"2025-06-06T11:20:24.771372Z","iopub.status.idle":"2025-06-06T11:20:24.787601Z","shell.execute_reply.started":"2025-06-06T11:20:24.771351Z","shell.execute_reply":"2025-06-06T11:20:24.786764Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def pick_rms(audio, window_size=512, top_db=15):\n    rms = librosa.feature.rms(y=audio, frame_length=window_size, hop_length=window_size).squeeze()\n    threshold = np.percentile(rms, 100 - top_db)\n    mask = rms >= threshold\n    mask_audio = np.repeat(mask, window_size)\n    mask_audio = mask_audio[:len(audio)]\n    return audio * mask_audio","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-06T11:20:24.788888Z","iopub.execute_input":"2025-06-06T11:20:24.789164Z","iopub.status.idle":"2025-06-06T11:20:24.802803Z","shell.execute_reply.started":"2025-06-06T11:20:24.789141Z","shell.execute_reply":"2025-06-06T11:20:24.802026Z"}},"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    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    audio_data = pick_rms(audio_data)\n    \n    mel_spec = audio2melspec(audio_data, cfg)\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)\n        \n    return mel_spec.astype(np.float32)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-06T11:20:24.803574Z","iopub.execute_input":"2025-06-06T11:20:24.803799Z","iopub.status.idle":"2025-06-06T11:20:24.815979Z","shell.execute_reply.started":"2025-06-06T11:20:24.803779Z","shell.execute_reply":"2025-06-06T11:20:24.815181Z"}},"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    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    for model_path in model_files:\n        try:\n            print(f\"Loading model: {model_path}\")\n            checkpoint = torch.load(model_path, map_location=torch.device(cfg.device))\n            \n            model = BirdCLEFModel(cfg, num_classes)\n            model.load_state_dict(checkpoint['model_state_dict'])\n            model = model.to(cfg.device)\n            model.eval()\n            \n            models.append(model)\n        except Exception as e:\n            print(f\"Error loading model {model_path}: {e}\")\n    \n    return models\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, using 10-sec padded input\"\"\"\n    predictions = []\n    row_ids = []\n    soundscape_id = Path(audio_path).stem\n\n    try:\n        print(f\"Processing {soundscape_id}\")\n        audio_segments = load_sample(audio_path, cfg)\n\n        for segment_idx, segment_audio in enumerate(audio_segments):\n            end_time_sec = (segment_idx + 1) * cfg.target_duration\n            row_id = f\"{soundscape_id}_{end_time_sec}\"\n            row_ids.append(row_id)\n\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 cfg.use_tta:\n                all_preds = []\n                for tta_idx in range(cfg.tta_count):\n                    mel_aug = apply_tta(mel_spec.squeeze(0).squeeze(0).numpy(), tta_idx)\n                    mel_aug = torch.tensor(mel_aug, dtype=torch.float32).unsqueeze(0).unsqueeze(0).to(cfg.device)\n\n                    if len(models) == 1:\n                        with torch.no_grad():\n                            outputs = models[0](mel_aug)\n                            probs = torch.sigmoid(outputs).cpu().numpy().squeeze()\n                            all_preds.append(probs)\n                    else:\n                        model_preds = []\n                        for model in models:\n                            with torch.no_grad():\n                                outputs = model(mel_aug)\n                                probs = torch.sigmoid(outputs).cpu().numpy().squeeze()\n                                model_preds.append(probs)\n                        all_preds.append(np.mean(model_preds, axis=0))\n                final_preds = np.mean(all_preds, axis=0)\n            else:\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                    model_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                            model_preds.append(probs)\n                    final_preds = np.mean(model_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    predictions = np.stack(predictions, axis=0)\n    return row_ids, predictions","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-06T11:20:24.816866Z","iopub.execute_input":"2025-06-06T11:20:24.817107Z","iopub.status.idle":"2025-06-06T11:20:24.831321Z","shell.execute_reply.started":"2025-06-06T11:20:24.817086Z","shell.execute_reply":"2025-06-06T11:20:24.830124Z"}},"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    test_files = list(Path(cfg.test_soundscapes).glob('*.ogg'))\n    \n    print(f\"Found {len(test_files)} test soundscapes\")\n\n    all_row_ids = []\n    all_predictions = []\n    grouped_preds = {}\n\n    for audio_path in tqdm(test_files):\n        row_ids, predictions = predict_on_spectrogram(str(audio_path), models, cfg, species_ids)\n        \n        predictions = np.array(predictions)\n        predictions = apply_power_to_low_ranked_cols(predictions, top_k=30, exponent=2.0)\n        smoothed_preds = predictions.copy()\n        for i in range(1, len(predictions) - 1):\n            smoothed_preds[i] = (\n                0.2 * predictions[i - 1] +\n                0.6 * predictions[i] +\n                0.2 * predictions[i + 1]\n            )\n        smoothed_preds[0] = 0.8 * predictions[0] + 0.2 * predictions[1]\n        smoothed_preds[-1] = 0.8 * predictions[-1] + 0.2 * predictions[-2]\n\n        all_row_ids.extend(row_ids)\n        all_predictions.extend(smoothed_preds)\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":"2025-06-06T11:20:24.832314Z","iopub.execute_input":"2025-06-06T11:20:24.832678Z","iopub.status.idle":"2025-06-06T11:20:24.848086Z","shell.execute_reply.started":"2025-06-06T11:20:24.832643Z","shell.execute_reply":"2025-06-06T11:20:24.847034Z"}},"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\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-06T11:20:24.849155Z","iopub.execute_input":"2025-06-06T11:20:24.849486Z","iopub.status.idle":"2025-06-06T11:20:24.865892Z","shell.execute_reply.started":"2025-06-06T11:20:24.849460Z","shell.execute_reply":"2025-06-06T11:20:24.864632Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"if __name__ == \"__main__\":\n    main()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-06T11:20:24.866971Z","iopub.execute_input":"2025-06-06T11:20:24.867362Z","iopub.status.idle":"2025-06-06T11:20:36.646991Z","shell.execute_reply.started":"2025-06-06T11:20:24.867337Z","shell.execute_reply":"2025-06-06T11:20:36.646205Z"}},"outputs":[],"execution_count":null}]}