{"metadata":{"kernelspec":{"display_name":"Python 3","language":"python","name":"python3"},"language_info":{"codemirror_mode":{"name":"ipython","version":3},"file_extension":".py","mimetype":"text/x-python","name":"python","nbconvert_exporter":"python","pygments_lexer":"ipython3","version":"3.10.12"},"papermill":{"default_parameters":{},"duration":4265.330291,"end_time":"2023-10-16T00:30:46.719691","environment_variables":{},"exception":null,"input_path":"__notebook__.ipynb","output_path":"__notebook__.ipynb","parameters":{},"start_time":"2023-10-15T23:19:41.389400","version":"2.4.0"},"kaggle":{"accelerator":"gpu","dataSources":[{"sourceId":21669,"databundleVersionId":1692278,"sourceType":"competition"}],"dockerImageVersionId":30588,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"import torch.nn as nn\nimport numpy as np\nimport torch\nimport librosa\nimport os\n\nfrom torch.utils.data import Dataset, DataLoader\nfrom sklearn.model_selection import KFold","metadata":{"_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","papermill":{"duration":5.503291,"end_time":"2023-10-15T23:19:49.800856","exception":false,"start_time":"2023-10-15T23:19:44.297565","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2023-11-26T11:17:09.153752Z","iopub.execute_input":"2023-11-26T11:17:09.154125Z","iopub.status.idle":"2023-11-26T11:17:13.282520Z","shell.execute_reply.started":"2023-11-26T11:17:09.154096Z","shell.execute_reply":"2023-11-26T11:17:13.281544Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"num_labels = 24\n\ndevice = 'cuda:0' if torch.cuda.is_available() else 'cpu'\nprint(device)","metadata":{"papermill":{"duration":0.069497,"end_time":"2023-10-15T23:19:49.873546","exception":false,"start_time":"2023-10-15T23:19:49.804049","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2023-11-26T11:19:43.354514Z","iopub.execute_input":"2023-11-26T11:19:43.354888Z","iopub.status.idle":"2023-11-26T11:19:43.360416Z","shell.execute_reply.started":"2023-11-26T11:19:43.354859Z","shell.execute_reply":"2023-11-26T11:19:43.359501Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Определяем функции для работы со спектрограммами","metadata":{}},{"cell_type":"code","source":"from skimage.transform import resize\nfrom skimage.filters import gaussian\nfrom skimage.color import rgb2gray\nfrom skimage import exposure, util\n\n\ndef horizontal_flip(img):\n    horizontal_flip_img = img[:, ::-1]\n    return addChannels(horizontal_flip_img)\n\ndef vertical_flip(img):\n    vertical_flip_img = img[::-1, :]\n    return addChannels(vertical_flip_img)\n\ndef addNoisy(img):\n    noise_img = util.random_noise(img)\n    return addChannels(noise_img)\n\ndef contrast_stretching(img):\n    contrast_img = exposure.rescale_intensity(img)\n    return addChannels(contrast_img)\n\ndef randomGaussian(img):\n    gaussian_img = gaussian(img)\n    return addChannels(gaussian_img)\n\ndef grayScale(img):\n    gray_img = rgb2gray(img)\n    return addChannels(gray_img)\n\ndef randomGamma(img):\n    img_gamma = exposure.adjust_gamma(img)\n    return addChannels(img_gamma)\n\ndef addChannels(img):\n    return np.stack((img, img, img))\n\ndef spec_to_image(spec):\n    spec = resize(spec, (224, 400))\n    eps=1e-6\n    mean = spec.mean()\n    std = spec.std()\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    return spec_scaled","metadata":{"papermill":{"duration":0.01267,"end_time":"2023-10-15T23:19:49.889036","exception":false,"start_time":"2023-10-15T23:19:49.876366","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2023-11-26T11:19:56.771362Z","iopub.execute_input":"2023-11-26T11:19:56.771729Z","iopub.status.idle":"2023-11-26T11:19:57.009269Z","shell.execute_reply.started":"2023-11-26T11:19:56.771702Z","shell.execute_reply":"2023-11-26T11:19:57.008479Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Считывание входных данных для обучения","metadata":{}},{"cell_type":"code","source":"import pandas as pd\n\nsr = 48000\nlength = 10 * sr\ndata = pd.read_csv(\"../input/rfcx-species-audio-detection/train_tp.csv\")\n\nfmin = sr / 2\nfmax = 0\nfor i in range(0, len(data)):\n    if fmin > float(data.iloc[i]['f_min']):\n        fmin = float(data.iloc[i]['f_min'])\n    if fmax < float(data.iloc[i]['f_max']):\n        fmax = float(data.iloc[i]['f_max'])\n        \nfmin = int(fmin * 0.9)\nfmax = int(fmax * 1.1)\n\nlabel_list = []\ndata_list = []\naudio_data = {}\nfor i in range(0, len(data)):\n    recording_id = data.recording_id.values[i]\n    species_id = int(data.species_id.values[i])\n    data_list.append(recording_id)\n    label_list.append(species_id)\n\n    wav, sr = librosa.load('../input/rfcx-species-audio-detection/train/' + recording_id + '.flac', sr=None)\n    t_min = float(data.t_min.values[i]) * sr\n    t_max = float(data.t_max.values[i]) * sr\n    center = np.round((t_min + t_max) / 2)\n    beginning = center - length / 2\n    if beginning < 0:\n        beginning = 0\n    ending = beginning + length\n    if ending > len(wav):\n        ending = len(wav)\n        beginning = ending - length\n    slice = wav[int(beginning):int(ending)]\n    \n    spec=librosa.feature.melspectrogram(y=slice, sr=sr, fmin=fmin, fmax=fmax)\n    spec_db=librosa.power_to_db(spec, top_db=80)\n    \n    img = spec_to_image(spec_db)\n    \n    audio_data[recording_id] = img","metadata":{"papermill":{"duration":283.122476,"end_time":"2023-10-15T23:24:33.028006","exception":false,"start_time":"2023-10-15T23:19:49.905530","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2023-11-26T11:58:00.686864Z","iopub.execute_input":"2023-11-26T11:58:00.687282Z","iopub.status.idle":"2023-11-26T11:59:20.709205Z","shell.execute_reply.started":"2023-11-26T11:58:00.687252Z","shell.execute_reply":"2023-11-26T11:59:20.707924Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Обучение","metadata":{}},{"cell_type":"code","source":"import copy\nfrom tqdm import tqdm\n\nlearning_rate = 1e-4\nepochs = 20\nloss_fn = nn.CrossEntropyLoss()\n\ndef train(model, loss_fn, train_loader, valid_loader, epochs, 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        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        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        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\" % (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    return model","metadata":{"execution":{"iopub.execute_input":"2023-10-15T23:24:33.076042Z","iopub.status.busy":"2023-10-15T23:24:33.075584Z","iopub.status.idle":"2023-10-15T23:24:33.094685Z","shell.execute_reply":"2023-10-15T23:24:33.093617Z"},"papermill":{"duration":0.030623,"end_time":"2023-10-15T23:24:33.097561","exception":false,"start_time":"2023-10-15T23:24:33.066938","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import random\n\nclass AudioData(Dataset):\n    def __init__(self, X, y, data_type):\n        self.data = []\n        self.labels = []\n        self.augs = [addNoisy, contrast_stretching,randomGaussian,randomGamma, vertical_flip, horizontal_flip, addChannels]\n        self.data_type=data_type\n        for i in range(0, len(X)):\n            recording_id = X[i]\n            label = y[i]\n            mel_spec = audio_data[recording_id]\n            self.data.append(mel_spec)\n            self.labels.append(label)\n                \n    def __len__(self):\n        return len(self.data)\n\n    def __getitem__(self, idx):\n        if self.data_type == \"train\":\n            aug= random.choice(self.augs)\n            data = aug(self.data[idx])\n        else:\n            data = addChannels(self.data[idx])\n        return data, self.labels[idx]","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"В качетсве модели была использована resnet101. Выбор ResNet-101 обусловлен его способностью извлекать сложные пространственные и временные зависимости из изображений, что является важным при анализе спектрограмм, представляющих звуковые данные.","metadata":{}},{"cell_type":"code","source":"from torchvision.models import resnet101\n\ndef get_model():\n    model = resnet101(pretrained=True)\n    num_ftrs = model.fc.in_features\n    model.fc = nn.Linear(num_ftrs, num_labels)\n    model = model.to(device)\n    return model","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"fold_num = 4\nskf = KFold(n_splits=fold_num, shuffle=True, random_state=32)\n\nfor fold_id, (train_index, val_index) in enumerate(skf.split(data_list, label_list)):\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\")\n    valid_data = AudioData(X_val, y_val, \"valid\")\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.AdamW(model.parameters(), lr=learning_rate)\n    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', patience=3)\n    model = train(model, loss_fn, train_loader, valid_loader, epochs, optimizer, scheduler)\n    torch.save(model.state_dict(), \"./model\" + str(fold_id) + \".pt\")\n    \n    del train_data, valid_data, train_loader, valid_loader, model, X_train, X_val, y_train, y_val","metadata":{"execution":{"iopub.execute_input":"2023-10-15T23:24:33.113603Z","iopub.status.busy":"2023-10-15T23:24:33.113142Z","iopub.status.idle":"2023-10-15T23:59:14.228421Z","shell.execute_reply":"2023-10-15T23:59:14.227463Z"},"papermill":{"duration":2081.125728,"end_time":"2023-10-15T23:59:14.230537","exception":false,"start_time":"2023-10-15T23:24:33.104809","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def load_test_file(f):\n    wav, sr = librosa.load('../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=fmin, fmax=fmax)\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":{"execution":{"iopub.execute_input":"2023-10-15T23:59:14.257013Z","iopub.status.busy":"2023-10-15T23:59:14.256693Z","iopub.status.idle":"2023-10-15T23:59:14.263294Z","shell.execute_reply":"2023-10-15T23:59:14.262368Z"},"papermill":{"duration":0.021679,"end_time":"2023-10-15T23:59:14.264999","exception":false,"start_time":"2023-10-15T23:59:14.243320","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"members = []\nfor i in range(fold_num):\n    model = get_model()\n    model.load_state_dict(torch.load('./model'+str(i)+'.pt'))\n    model.eval()\n    members.append(model)\n    \nos.remove('./model0.pt') \nos.remove('./model1.pt')\nos.remove('./model2.pt') \nos.remove('./model3.pt')","metadata":{"execution":{"iopub.execute_input":"2023-10-15T23:59:14.292167Z","iopub.status.busy":"2023-10-15T23:59:14.291923Z","iopub.status.idle":"2023-10-15T23:59:19.576367Z","shell.execute_reply":"2023-10-15T23:59:19.573855Z"},"papermill":{"duration":5.303173,"end_time":"2023-10-15T23:59:19.580107","exception":false,"start_time":"2023-10-15T23:59:14.276934","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import csv\n\nwith open('submission.csv', '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','s11',\n                               's12','s13','s14','s15','s16','s17','s18','s19','s20','s21','s22','s23'])\n    \n    test_files = os.listdir('../input/rfcx-species-audio-detection/test/')\n    print(len(test_files))\n    \n    for i in range(0, len(test_files)):\n        data = load_test_file(test_files[i])\n        data = torch.tensor(data)\n        data = data.float()\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        avg_maxed_output = torch.mean(torch.stack(output_list), dim=0)\n        \n        file_id = str.split(test_files[i], '.')[0]\n        write_array = [file_id]\n        \n        for out in avg_maxed_output:\n            write_array.append(out.item())\n    \n        submission_writer.writerow(write_array)\n","metadata":{"execution":{"iopub.execute_input":"2023-10-15T23:59:19.849469Z","iopub.status.busy":"2023-10-15T23:59:19.848923Z","iopub.status.idle":"2023-10-16T00:30:44.063842Z","shell.execute_reply":"2023-10-16T00:30:44.062940Z"},"papermill":{"duration":1884.243432,"end_time":"2023-10-16T00:30:44.079646","exception":false,"start_time":"2023-10-15T23:59:19.836214","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]}]}