{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.11.13","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"nvidiaTeslaT4","dataSources":[{"sourceId":21669,"databundleVersionId":1692278,"sourceType":"competition"}],"dockerImageVersionId":31192,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"import torch.nn as nn\nimport pandas as pd\nimport numpy as np\nimport librosa\nimport random\nimport torch\nimport copy\nimport csv\nimport os\n\nfrom concurrent.futures import ThreadPoolExecutor\nfrom torch.utils.data import Dataset, DataLoader\nfrom sklearn.model_selection import KFold\nfrom torchvision.models import resnet50\nfrom skimage.filters import gaussian\nfrom skimage.transform import resize\nfrom skimage import exposure, util\nfrom tqdm import tqdm\n\nimport warnings\nwarnings.filterwarnings('ignore')\n\ndevice = 'cuda:0' if torch.cuda.is_available() else 'cpu'\nprint(device)\n\n\nLABELS = 24\nSR = 48000\nLENGTH = 10 * SR\nF_MIN = 24000\nF_MAX = 0\nLEARNING_RATE = 2e-4\nEPOCHS = 20\nN_FOLD = 5","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-12-13T17:56:40.371461Z","iopub.execute_input":"2025-12-13T17:56:40.372136Z","iopub.status.idle":"2025-12-13T17:56:52.672979Z","shell.execute_reply.started":"2025-12-13T17:56:40.372108Z","shell.execute_reply":"2025-12-13T17:56:52.672314Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class AudioAugmentations:\n    def __init__(self):\n        self.augs = [self.add_noise, self.contrast_stretch, self.h_flip, self.v_flip]\n\n    def h_flip(self, image):\n        return np.stack([image[:, ::-1]] * 3)\n\n    def v_flip(self, image):\n        return np.stack([image[::-1, :]] * 3)\n\n    def add_noise(self, image):\n        noise_img = util.random_noise(image)\n        return np.stack([noise_img] * 3)\n\n    def contrast_stretch(self, image):\n        contrast_img = exposure.rescale_intensity(image)\n        return np.stack([contrast_img] * 3)\n\n    def apply_random_augmentation(self, image):\n        aug_func = random.choice(self.augs)\n        return aug_func(image)","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def spec_to_image(spec):\n    spec = resize(spec, (224, 400))\n    eps=1e-6\n\n    mean = spec.mean()\n    std = spec.std()\n\n    spec_norm = (spec - mean) / (std + eps)\n    spec_min, spec_max = spec_norm.min(), spec_norm.max()\n    spec_scaled = 255 * (spec_norm - spec_min) / (spec_max - spec_min)\n    spec_scaled = spec_scaled.astype(np.uint8)\n    spec_scaled = np.asarray(spec_scaled)\n\n    return spec_scaled","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def get_model():\n    model = resnet50(pretrained=True)\n    num_ftrs = model.fc.in_features\n    model.fc = nn.Linear(num_ftrs, LABELS)\n\n    return model.to(device)","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"data = pd.read_csv(\"/kaggle/input/rfcx-species-audio-detection/train_tp.csv\")\n\nfor i in range(0, len(data)):\n    if F_MIN > float(data.iloc[i]['f_min']):\n        F_MIN = float(data.iloc[i]['f_min'])\n    if F_MAX < float(data.iloc[i]['f_max']):\n        F_MAX = float(data.iloc[i]['f_max'])\n\nF_MIN = int(F_MIN * 0.9)\nF_MAX = int(F_MAX * 1.1)","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"label_list = data['species_id'].tolist()\ndata_list = data['recording_id'].tolist()\naudio_data = {}\n\ndef process_audio(i):\n    recording_id = data_list[i]\n    species_id = label_list[i]\n\n    wav, sr = librosa.load(f'/kaggle/input/rfcx-species-audio-detection/train/{recording_id}.flac', sr=None)\n\n    t_min = int(data.at[i, 't_min'] * sr)\n    t_max = int(data.at[i, 't_max'] * sr)\n\n    center = np.round((t_min + t_max) / 2)\n    beginning = max(center - LENGTH // 2, 0)\n    ending = min(beginning + LENGTH, len(wav))\n\n    beginning = ending - LENGTH if ending - beginning < LENGTH else beginning\n    slice = wav[int(beginning):int(ending)]\n    spec = librosa.feature.melspectrogram(y=slice, sr=sr, fmin=F_MIN, fmax=F_MAX)\n    spec_db = librosa.power_to_db(spec, top_db=80)\n\n    img = spec_to_image(spec_db)\n\n    return recording_id, img","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"with ThreadPoolExecutor() as executor:\n    results = list(executor.map(process_audio, range(len(data))))\n\nfor recording_id, img in results:\n    audio_data[recording_id] = img","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class AudioData(Dataset):\n    def __init__(self, X, y, data_type, augmentations=None):\n        self.X = X\n        self.y = y\n        self.data_type = data_type\n        self.audio_data = audio_data\n        self.augmentations = augmentations\n\n    def __len__(self):\n        return len(self.X)\n\n    def __getitem__(self, idx):\n        recording_id = self.X[idx]\n        label = self.y[idx]\n\n        img = self.audio_data[recording_id]\n\n        if self.data_type == \"train\" and self.augmentations:\n            img = self.augmentations.apply_random_augmentation(img)\n        else:\n            img = np.stack((img, img, img))\n\n        return img, label","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"loss_fn = nn.CrossEntropyLoss()\naudio_augmenter = AudioAugmentations()\n\ndef train(model, loss_fn, train_loader, valid_loader, optimizer, scheduler):\n    best_model_wts = copy.deepcopy(model.state_dict())\n    best_acc = 0.0\n    train_losses = []\n    valid_losses = []\n\n    for epoch in tqdm(range(1, EPOCHS + 1)):\n        model.train()\n        batch_losses = []\n\n        for _, data in enumerate(train_loader):\n            x, y = data\n            optimizer.zero_grad()\n            x = x.to(device, dtype=torch.float32)\n            y = y.to(device, dtype=torch.long)\n            y_hat = model(x)\n            loss = loss_fn(y_hat, y)\n            loss.backward()\n            batch_losses.append(loss.item())\n            optimizer.step()\n        train_losses.append(batch_losses)\n\n        model.eval()\n        batch_losses = []\n        trace_y = []\n        trace_yhat = []\n\n        with torch.no_grad():\n            for _, data in enumerate(valid_loader):\n                x, y = data\n                x = x.to(device, dtype=torch.float32)\n                y = y.to(device, dtype=torch.long)\n                y_hat = model(x)\n                loss = loss_fn(y_hat, y)\n                trace_y.append(y.cpu().detach().numpy())\n                trace_yhat.append(y_hat.cpu().detach().numpy())\n                batch_losses.append(loss.item())\n\n        valid_losses.append(batch_losses)\n        trace_y = np.concatenate(trace_y)\n        trace_yhat = np.concatenate(trace_yhat)\n        accuracy = np.mean(trace_yhat.argmax(axis=1) == trace_y)\n\n        print(\"epoch = %d, train_loss = %.5f, val_loss = %.5f, val_accuracy = %.5f\" % (\n            epoch, np.mean(train_losses[-1]), np.mean(valid_losses[-1]), accuracy))\n\n        scheduler.step(np.mean(valid_losses[-1]))\n        if accuracy > best_acc:\n            best_acc = accuracy\n            best_model_wts = copy.deepcopy(model.state_dict())\n\n    model.load_state_dict(best_model_wts)\n\n    return model","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"skf = KFold(n_splits=N_FOLD, shuffle=True, random_state=563)\n\nfor fold_id, (train_index, val_index) in enumerate(skf.split(data_list, label_list)):\n    print(\"Fold\", fold_id)\n\n    X_train = np.take(data_list, train_index)\n    y_train = np.take(label_list, train_index, axis=0)\n    X_val = np.take(data_list, val_index)\n    y_val = np.take(label_list, val_index, axis=0)\n\n    train_data = AudioData(X_train, y_train, \"train\", augmentations=audio_augmenter)\n    valid_data = AudioData(X_val, y_val, \"valid\")\n\n    train_loader = DataLoader(train_data, batch_size=8, shuffle=True, drop_last=True)\n    valid_loader = DataLoader(valid_data, batch_size=8, shuffle=True, drop_last=True)\n\n    model = get_model()\n    optimizer = torch.optim.Adam(model.parameters(), lr=LEARNING_RATE)\n    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', patience=3)\n\n    model = train(model, loss_fn, train_loader, valid_loader, optimizer, scheduler)\n    torch.save(model.state_dict(), f\"./model{fold_id}.pt\")\n\n    del train_data, valid_data, train_loader, valid_loader, model, X_train, X_val, y_train, y_val","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def load_test_file(f):\n    wav, sr = librosa.load('/kaggle/input/rfcx-species-audio-detection/test/' + f, sr=None)\n\n    segments = len(wav) / LENGTH\n    segments = int(np.ceil(segments))\n\n    mel_array = []\n\n    for i in range(0, segments):\n        if (i + 1) * LENGTH > len(wav):\n            slice = wav[len(wav) - LENGTH:len(wav)]\n        else:\n            slice = wav[i * LENGTH:(i + 1) * LENGTH]\n\n        spec = librosa.feature.melspectrogram(y=slice, sr=sr, fmin=F_MIN, fmax=F_MAX)\n        spec_db = librosa.power_to_db(spec, top_db=80)\n\n        img = spec_to_image(spec_db)\n        mel_spec = np.stack((img, img, img))\n        mel_array.append(mel_spec)\n\n    return mel_array","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"members = []\n\nfor i in range(N_FOLD):\n    model = get_model()\n\n    model.load_state_dict(torch.load('./model' + str(i) + '.pt'))\n    model.eval()\n\n    members.append(model)\n\n    os.remove('./model' + str(i) + '.pt')","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def load_and_predict(test_file, members):\n    data = load_test_file(test_file)\n    data = torch.tensor(data).float()\n\n    if torch.cuda.is_available():\n        data = data.cuda()\n\n    output_list = []\n    for m in members:\n        output = m(data)\n        maxed_output = torch.max(output, dim=0)[0]\n        maxed_output = maxed_output.cpu().detach()\n        output_list.append(maxed_output)\n\n    avg_maxed_output = torch.mean(torch.stack(output_list), dim=0)\n    file_id = test_file.split('.')[0]\n    return [file_id] + [out.item() for out in avg_maxed_output]\n\ndef save_submission(predictions, output_file='submission.csv'):\n    with open(output_file, 'w', newline='') as csvfile:\n        submission_writer = csv.writer(csvfile, delimiter=',')\n        submission_writer.writerow(['recording_id', 's0', 's1', 's2', 's3', 's4', 's5', 's6', 's7', 's8', 's9', 's10', \n                                    's11', 's12', 's13', 's14', 's15', 's16', 's17', 's18', 's19', 's20', 's21', 's22', 's23'])\n        for pred in predictions:\n            submission_writer.writerow(pred)\n\ndef generate_predictions(test_files, members):\n    predictions = []\n\n    with ThreadPoolExecutor(max_workers=4) as executor:\n        futures = [executor.submit(load_and_predict, test_file, members) for test_file in test_files]\n        for future in futures:\n            predictions.append(future.result())\n\n    save_submission(predictions)","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"test_files = os.listdir('/kaggle/input/rfcx-species-audio-detection/test/')\n\n\nif torch.cuda.is_available():\n    members = [m.cuda() for m in members]\n\ngenerate_predictions(test_files, members)","metadata":{"trusted":true},"outputs":[],"execution_count":null}]}