{"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,"isSourceIdPinned":false,"sourceType":"competition"},{"sourceId":422869,"sourceType":"modelInstanceVersion","isSourceIdPinned":false,"modelInstanceId":344597,"modelId":365892},{"sourceId":426155,"sourceType":"modelInstanceVersion","isSourceIdPinned":false,"modelInstanceId":347396,"modelId":368655}],"dockerImageVersionId":30918,"isInternetEnabled":false,"language":"python","sourceType":"notebook","isGpuEnabled":false}},"nbformat_minor":4,"nbformat":4,"cells":[{"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":"2025-06-05T20:45:34.335077Z","iopub.execute_input":"2025-06-05T20:45:34.335418Z","iopub.status.idle":"2025-06-05T20:45:34.341281Z","shell.execute_reply.started":"2025-06-05T20:45:34.335391Z","shell.execute_reply":"2025-06-05T20:45:34.340033Z"}},"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_paths = [\n    ('/kaggle/input/effnet_if/pytorch/default/1', 'efficientnet_b0'),\n    ('/kaggle/input/effnet_if/pytorch/default/1', 'efficientnet_b2'),\n    ('/kaggle/input/regnet_5fold/pytorch/default/1', 'regnety_008')\n    ]\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    \n    in_channels = 1\n    device = 'cpu'  \n    \n    # Inference parameters\n    batch_size = 16\n    use_tta = True  \n    tta_count = 2   \n    threshold = 0.5\n    \n    use_specific_folds = True  # If False, use all found models\n    folds = [ 2, 4]  # Used only if use_specific_folds is True\n    \n    debug = False\n    debug_count = 10\n\ncfg = CFG()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-05T20:57:29.930440Z","iopub.execute_input":"2025-06-05T20:57:29.930931Z","iopub.status.idle":"2025-06-05T20:57:29.939364Z","shell.execute_reply.started":"2025-06-05T20:57:29.930885Z","shell.execute_reply":"2025-06-05T20:57:29.937898Z"}},"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":"2025-06-05T20:45:34.365368Z","iopub.execute_input":"2025-06-05T20:45:34.365785Z","iopub.status.idle":"2025-06-05T20:45:34.391195Z","shell.execute_reply.started":"2025-06-05T20:45:34.365752Z","shell.execute_reply":"2025-06-05T20:45:34.389958Z"}},"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        if 'efficientnet' in cfg.model_name:\n            self.backbone = timm.create_model(\n            cfg.model_name,\n            pretrained=False,  \n            in_chans=cfg.in_channels,\n            features_only=True,\n            out_indices=(3, 4),\n            drop_rate=0.0,    \n            drop_path_rate=0.0\n        )\n            self.pooling = nn.AdaptiveAvgPool2d(1)\n            \n            selected_indices = (3, 4)\n            self.feat_dim = sum([self.backbone.feature_info[i]['num_chs'] for i in selected_indices])\n            self.classifier = nn.Linear(self.feat_dim, num_classes)\n            \n        else:\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            backbone_out = self.backbone.get_classifier().in_features\n            self.backbone.reset_classifier(0, '')\n            self.pooling = nn.AdaptiveAvgPool2d(1)\n            self.feat_dim = backbone_out\n            self.classifier = nn.Linear(backbone_out, num_classes)\n        \n        \n    def forward(self, x):\n        features = self.backbone(x)\n\n        if 'efficientnet' in self.cfg.model_name:\n            pooled = [self.pooling(f) for f in features]\n        \n            # Flatten and concatenate\n            pooled = [p.view(p.size(0), -1) for p in pooled]\n            features = torch.cat(pooled, dim=1)\n        else:\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":"2025-06-05T20:54:49.025290Z","iopub.execute_input":"2025-06-05T20:54:49.025772Z","iopub.status.idle":"2025-06-05T20:54:49.041362Z","shell.execute_reply.started":"2025-06-05T20:54:49.025731Z","shell.execute_reply":"2025-06-05T20:54:49.040084Z"}},"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    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":"2025-06-05T20:45:34.411790Z","iopub.execute_input":"2025-06-05T20:45:34.412124Z","iopub.status.idle":"2025-06-05T20:45:34.436676Z","shell.execute_reply.started":"2025-06-05T20:45:34.412092Z","shell.execute_reply":"2025-06-05T20:45:34.435410Z"}},"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    for model_dir, model_name in cfg.model_paths:\n        model_files = [str(p) for p in Path(model_dir).rglob(\"*.pth\")]\n\n        if not model_files:\n            print(f\"Warning: No model files found under {model_dir}!\")\n            continue\n\n        if model_name == 'efficientnet_b2':\n            model_files = [f for f in model_files if \"b2\" in f]\n            print(f\"EfficientNet-B2 → loading fold1 only: {len(model_files)} file(s).\")\n\n        elif model_name == 'efficientnet_b0':\n            model_files = [f for f in model_files if \"b0\" in f]\n            print(f\"EfficientNet-B0 → loading fold1 only: {len(model_files)} file(s).\")\n\n        elif model_name == 'regnety_008':\n            filtered = []\n            for fold in [2, 4]:\n                filtered.extend([f for f in model_files if f\"fold{fold}\" in f])\n            model_files = filtered\n            print(f\"RegNetY_008 → loading folds [2, 4]: {len(model_files)} file(s).\")\n\n        else:\n            if cfg.use_specific_folds:\n                filtered = []\n                for fold in cfg.folds:\n                    filtered.extend([f for f in model_files if f\"fold{fold}\" in f])\n                model_files = filtered\n                print(f\"{model_name} → loading folds {cfg.folds}: {len(model_files)} file(s).\")\n            else:\n                print(f\"{model_name} → loading all {len(model_files)} file(s).\")\n\n        for ckpt_path in model_files:\n            try:\n                print(f\"Loading model: {ckpt_path}\")\n                checkpoint = torch.load(ckpt_path, map_location=torch.device(cfg.device))\n\n                temp_cfg = CFG()\n                temp_cfg.model_name = model_name\n                temp_cfg.in_channels = cfg.in_channels\n                temp_cfg.device = cfg.device\n\n                model = BirdCLEFModel(temp_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 {ckpt_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\"\"\"\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                            outputs = models[0](mel_spec)\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                                outputs = model(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":"2025-06-05T20:48:19.973859Z","iopub.execute_input":"2025-06-05T20:48:19.974270Z","iopub.status.idle":"2025-06-05T20:48:19.995671Z","shell.execute_reply.started":"2025-06-05T20:48:19.974238Z","shell.execute_reply":"2025-06-05T20:48:19.994592Z"}},"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).copy()\n    elif tta_idx == 2:\n        # Frequency shift (vertical flip)\n        return np.flip(spec, axis=0).copy()\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\n    submission_dict = {'row_id': row_ids}\n    #small testing change - species_ids\n    test=['bubwre1']\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":"2025-06-05T20:45:34.481495Z","iopub.execute_input":"2025-06-05T20:45:34.481974Z","iopub.status.idle":"2025-06-05T20:45:34.510953Z","shell.execute_reply.started":"2025-06-05T20:45:34.481905Z","shell.execute_reply":"2025-06-05T20:45:34.509693Z"}},"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\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":"2025-06-05T20:45:34.512957Z","iopub.execute_input":"2025-06-05T20:45:34.513362Z","iopub.status.idle":"2025-06-05T20:45:34.534640Z","shell.execute_reply.started":"2025-06-05T20:45:34.513320Z","shell.execute_reply":"2025-06-05T20:45:34.533645Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"if __name__ == \"__main__\":\n    main()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-05T20:54:55.025784Z","iopub.execute_input":"2025-06-05T20:54:55.026120Z","iopub.status.idle":"2025-06-05T20:55:24.722413Z","shell.execute_reply.started":"2025-06-05T20:54:55.026093Z","shell.execute_reply":"2025-06-05T20:55:24.721246Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null}]}