{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"pygments_lexer":"ipython3","nbconvert_exporter":"python","version":"3.6.4","file_extension":".py","codemirror_mode":{"name":"ipython","version":3},"name":"python","mimetype":"text/x-python"}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"! pip3 install timm -q","metadata":{"id":"uBZokyT1gYWD","outputId":"aed15e9e-e33e-4780-9434-cf85f12c2229","execution":{"iopub.status.busy":"2023-02-28T04:18:59.477627Z","iopub.execute_input":"2023-02-28T04:18:59.478203Z","iopub.status.idle":"2023-02-28T04:19:11.243948Z","shell.execute_reply.started":"2023-02-28T04:18:59.478154Z","shell.execute_reply":"2023-02-28T04:19:11.24269Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nimport time\nimport h5py\nimport timm\nimport torch\nimport torch.nn as nn\nimport torchaudio\nimport torchvision.transforms as TF\nimport seaborn as sns\nsns.set_theme(style='dark')\n\n\nfrom tqdm.auto import tqdm\nfrom sklearn.model_selection import KFold\nfrom sklearn.metrics import roc_auc_score\nfrom timm.scheduler import CosineLRScheduler\n\ndevice = torch.device('cuda')\ncriterion = nn.BCEWithLogitsLoss()\n\n# метаданные train\ndi = '../input/g2net-detecting-continuous-gravitational-waves'\ndf = pd.read_csv(di + '/train_labels.csv')\ndf = df[df.target >= 0]  # удаление меток -1","metadata":{"id":"3Xgw6q0ugYWH","execution":{"iopub.status.busy":"2023-02-28T04:19:11.24663Z","iopub.execute_input":"2023-02-28T04:19:11.247148Z","iopub.status.idle":"2023-02-28T04:19:11.286328Z","shell.execute_reply.started":"2023-02-28T04:19:11.247106Z","shell.execute_reply":"2023-02-28T04:19:11.285481Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Класс Dataset и параметры","metadata":{"id":"tbekTVh-gYWI"}},{"cell_type":"code","source":"transforms_time_mask = nn.Sequential(\n                torchaudio.transforms.TimeMasking(time_mask_param=10),\n            )\n\ntransforms_freq_mask = nn.Sequential(\n                torchaudio.transforms.FrequencyMasking(freq_mask_param=10),\n            )\n\nflip_rate = 0.0 # вероятность применения горизонтального и вертикального переворота\nfre_shift_rate = 0.0 # вероятность применения вертикального сдвига\n\ntime_mask_num = 0 # количество масок по времени\nfreq_mask_num = 0 # количество масок по частоте","metadata":{"id":"7Z3rNynhq1gr","execution":{"iopub.status.busy":"2023-02-28T04:19:11.288046Z","iopub.execute_input":"2023-02-28T04:19:11.288416Z","iopub.status.idle":"2023-02-28T04:19:11.295109Z","shell.execute_reply.started":"2023-02-28T04:19:11.288371Z","shell.execute_reply":"2023-02-28T04:19:11.293867Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class Dataset(torch.utils.data.Dataset):\n    \"\"\"\n    dataset = Dataset(data_type, df)\n\n    img, y = dataset[i]\n      img (np.float32): 2 x 360 x 128\n      y (np.float32): метка 0 или 1\n    \"\"\"\n    def __init__(self, data_type, df, tfms=False):\n        self.data_type = data_type\n        self.df = df\n        self.tfms = tfms\n\n    def __len__(self):\n        return len(self.df)\n\n    def __getitem__(self, i):\n        \"\"\"\n        i (int): получение i-ого элемента\n        \"\"\"\n        r = self.df.iloc[i]\n        y = np.float32(r.target)\n        file_id = r.id\n\n        img = np.empty((2, 360, 128), dtype=np.float32)\n\n        filename = '%s/%s/%s.hdf5' % (di, self.data_type, file_id)\n        with h5py.File(filename, 'r') as f:\n            g = f[file_id]\n\n            for ch, s in enumerate(['H1', 'L1']):\n                a = g[s]['SFTs'][:, :4096] * 1e22  # обратное преобразование Фурье\n\n                p = a.real**2 + a.imag**2\n                p /= np.mean(p)\n                p = np.mean(p.reshape(360, 128, 32), axis=2)  # 4096 -> 128\n                img[ch] = p\n\n        if self.tfms:\n            if np.random.rand() <= flip_rate: # горизонтальный переворот\n                img = np.flip(img, axis=1).copy()\n            if np.random.rand() <= flip_rate: # вертикальный переворот\n                img = np.flip(img, axis=2).copy()\n            if np.random.rand() <= fre_shift_rate: # вертикальный сдвиг\n                img = np.roll(img, np.random.randint(low=0, high=img.shape[1]), axis=1)\n            \n            img = torch.from_numpy(img)\n\n            for _ in range(time_mask_num): # маска по времени\n                img = transforms_time_mask(img)\n            for _ in range(freq_mask_num): # маска по частоте\n                img = transforms_freq_mask(img)\n        \n        else:\n            img = torch.from_numpy(img)\n                \n        return img, y","metadata":{"id":"3kOMMsyagYWN","execution":{"iopub.status.busy":"2023-02-28T04:19:11.296861Z","iopub.execute_input":"2023-02-28T04:19:11.297523Z","iopub.status.idle":"2023-02-28T04:19:11.312903Z","shell.execute_reply.started":"2023-02-28T04:19:11.297488Z","shell.execute_reply":"2023-02-28T04:19:11.311904Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Аугментация данных","metadata":{}},{"cell_type":"markdown","source":"## Горизонтальный и вертикальный переворот","metadata":{}},{"cell_type":"code","source":"dataset = Dataset('train', df, tfms=False)\nimg, y = dataset[10]\n\nplt.subplot(1, 2, 1)\nplt.title('Спектрограмма')\nplt.xlabel('Время')\nplt.ylabel('Частота')\nplt.imshow(img[0, 0:360])\nplt.colorbar()\n\n\n\nflip_rate = 1.0\n\ndataset = Dataset('train', df, tfms=True)\nimg, y = dataset[10]\n\nplt.subplot(1, 2, 2)\nplt.title('Спектрограмма')\nplt.xlabel('Время')\nplt.ylabel('Частота')\nplt.imshow(img[0, 0:360])\nplt.colorbar()\n\nplt.show()","metadata":{"execution":{"iopub.status.busy":"2023-02-28T04:19:11.316228Z","iopub.execute_input":"2023-02-28T04:19:11.316622Z","iopub.status.idle":"2023-02-28T04:19:12.582781Z","shell.execute_reply.started":"2023-02-28T04:19:11.316596Z","shell.execute_reply":"2023-02-28T04:19:12.581798Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Вертикальный сдвиг","metadata":{}},{"cell_type":"code","source":"dataset = Dataset('train', df, tfms=False)\nimg, y = dataset[10]\n\n\nplt.subplot(1, 2, 1)\nplt.title('Спектрограмма')\nplt.xlabel('Время')\nplt.ylabel('Частота')\nplt.imshow(img[0, 0:360])\nplt.colorbar()\n\n\n\nflip_rate = 0.0\nfre_shift_rate = 1.0\n\ndataset = Dataset('train', df, tfms=True)\nimg, y = dataset[10]\n\nplt.subplot(1, 2, 2)\nplt.title('Спектрограмма')\nplt.xlabel('Время')\nplt.ylabel('Частота')\nplt.imshow(img[0, 0:360])\nplt.colorbar()\n\nplt.show()","metadata":{"id":"KKX7AmjTU8qI","outputId":"448f6ad0-bae2-4e83-8028-c86fdd2fa17b","execution":{"iopub.status.busy":"2023-02-28T04:19:12.584165Z","iopub.execute_input":"2023-02-28T04:19:12.585157Z","iopub.status.idle":"2023-02-28T04:19:13.600496Z","shell.execute_reply.started":"2023-02-28T04:19:12.585122Z","shell.execute_reply":"2023-02-28T04:19:13.599514Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Маска по времени","metadata":{}},{"cell_type":"code","source":"dataset = Dataset('train', df, tfms=False)\nimg, y = dataset[10]\n\n\nplt.subplot(1, 2, 1)\nplt.title('Спектрограмма')\nplt.xlabel('Время')\nplt.ylabel('Частота')\nplt.imshow(img[0, 0:360])\nplt.colorbar()\n\n\nflip_rate = 0.0\nfre_shift_rate = 0.0\ntime_mask_num = 3\n\ndataset = Dataset('train', df, tfms=True)\nimg, y = dataset[10]\n\nplt.subplot(1, 2, 2)\nplt.title('Спектрограмма')\nplt.xlabel('Время')\nplt.ylabel('Частота')\nplt.imshow(img[0, 0:360])\nplt.colorbar()\n\nplt.show()","metadata":{"execution":{"iopub.status.busy":"2023-02-28T04:19:13.601936Z","iopub.execute_input":"2023-02-28T04:19:13.602784Z","iopub.status.idle":"2023-02-28T04:19:14.620662Z","shell.execute_reply.started":"2023-02-28T04:19:13.602747Z","shell.execute_reply":"2023-02-28T04:19:14.619785Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Маска по частоте","metadata":{}},{"cell_type":"code","source":"dataset = Dataset('train', df, tfms=False)\nimg, y = dataset[10]\n\n\nplt.subplot(1, 2, 1)\nplt.title('Спектрограмма')\nplt.xlabel('Время')\nplt.ylabel('Частота')\nplt.imshow(img[0, 0:360])\nplt.colorbar()\n\n\nflip_rate = 0.0\nfre_shift_rate = 0.0\ntime_mask_num = 0\nfreq_mask_num = 3\n\ndataset = Dataset('train', df, tfms=True)\nimg, y = dataset[10]\n\nplt.subplot(1, 2, 2)\nplt.title('Спектрограмма')\nplt.xlabel('Время')\nplt.ylabel('Частота')\nplt.imshow(img[0, 0:360])\nplt.colorbar()\n\nplt.show()","metadata":{"execution":{"iopub.status.busy":"2023-02-28T04:19:14.622053Z","iopub.execute_input":"2023-02-28T04:19:14.623089Z","iopub.status.idle":"2023-02-28T04:19:15.658824Z","shell.execute_reply.started":"2023-02-28T04:19:14.623044Z","shell.execute_reply":"2023-02-28T04:19:15.657906Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Модель","metadata":{"id":"ejqEYZTxgYWP"}},{"cell_type":"code","source":"class Model(nn.Module):\n    def __init__(self, name, *, pretrained=False):\n        \"\"\"\n        name (str): название модели из timm\n        \"\"\"\n        super().__init__()\n\n        # Use timm\n        model = timm.create_model(name, pretrained=pretrained, in_chans=2)\n\n        clsf = model.default_cfg['classifier']\n        n_features = model._modules[clsf].in_features\n        model._modules[clsf] = nn.Identity()\n\n        self.fc = nn.Linear(n_features, 1)\n        self.model = model\n\n    def forward(self, x):\n        x = self.model(x)\n        x = self.fc(x)\n        return x","metadata":{"id":"W3JQE_UbgYWP","execution":{"iopub.status.busy":"2023-02-28T04:19:15.660473Z","iopub.execute_input":"2023-02-28T04:19:15.664383Z","iopub.status.idle":"2023-02-28T04:19:15.67403Z","shell.execute_reply.started":"2023-02-28T04:19:15.66434Z","shell.execute_reply":"2023-02-28T04:19:15.672274Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Предсказание и оценка","metadata":{"id":"eO6YqnT0gYWQ"}},{"cell_type":"code","source":"def evaluate(model, loader_val, *, compute_score=True, pbar=None):\n    \"\"\"\n    Предсказание и расчет loss и score\n    \"\"\"\n    tb = time.time()\n    was_training = model.training\n    model.eval()\n\n    loss_sum = 0.0\n    n_sum = 0\n    y_all = []\n    y_pred_all = []\n\n    if pbar is not None:\n        pbar = tqdm(desc='Predict', nrows=78, total=pbar)\n\n    for img, y in loader_val:\n        n = y.size(0)\n        img = img.to(device)\n        y = y.to(device)\n\n        with torch.no_grad():\n                y_pred = model(img.to(device))\n\n        loss = criterion(y_pred.view(-1), y)\n\n        n_sum += n\n        loss_sum += n * loss.item()\n\n        y_all.append(y.cpu().detach().numpy())\n        y_pred_all.append(y_pred.sigmoid().squeeze().cpu().detach().numpy())\n\n        if pbar is not None:\n            pbar.update(len(img))\n        \n        del loss, y_pred, img, y\n\n    loss_val = loss_sum / n_sum\n\n    y = np.concatenate(y_all)\n    y_pred = np.concatenate(y_pred_all)\n\n    score = roc_auc_score(y, y_pred) if compute_score else None\n\n    ret = {'loss': loss_val,\n           'score': score,\n           'y': y,\n           'y_pred': y_pred,\n           'time': time.time() - tb}\n    \n    model.train(was_training)\n\n    return ret","metadata":{"id":"B-xWfjWYgYWR","execution":{"iopub.status.busy":"2023-02-28T04:19:15.675711Z","iopub.execute_input":"2023-02-28T04:19:15.676546Z","iopub.status.idle":"2023-02-28T04:19:15.697802Z","shell.execute_reply.started":"2023-02-28T04:19:15.6764Z","shell.execute_reply":"2023-02-28T04:19:15.696949Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Обучение","metadata":{"id":"3arVeLkRgYWS"}},{"cell_type":"code","source":"model_name = 'tf_efficientnet_b6_ns'\nnfold = 5\nkfold = KFold(n_splits=nfold, random_state=42, shuffle=True)\n\nepochs = 25\nbatch_size = 32\nnum_workers = 2\nweight_decay = 1e-6\nmax_grad_norm = 1000\n\nlr_max = 4e-4\nepochs_warmup = 1.0\n\n\n## настройки аугментаций\nflip_rate = 0.5\nfre_shift_rate = 1.0\ntime_mask_num = 1\nfreq_mask_num = 2\n\nfor ifold, (idx_train, idx_test) in enumerate(kfold.split(df)):\n    print('Fold %d/%d' % (ifold + 1, nfold))\n    torch.manual_seed(42 + ifold + 1)\n\n    # Train - val split\n    dataset_train = Dataset('train', df.iloc[idx_train], tfms=True)\n    dataset_val = Dataset('train', df.iloc[idx_test])\n\n    loader_train = torch.utils.data.DataLoader(dataset_train, batch_size=batch_size,\n                     num_workers=num_workers, pin_memory=True, shuffle=True, drop_last=True)\n    loader_val = torch.utils.data.DataLoader(dataset_val, batch_size=batch_size,\n                     num_workers=num_workers, pin_memory=True)\n\n    # Модель и оптимизатор\n    model = Model(model_name, pretrained=True)\n    model.to(device)\n    model.train()\n\n    optimizer = torch.optim.AdamW(model.parameters(), lr=lr_max, weight_decay=weight_decay)\n\n    # График Learning-rate\n    nbatch = len(loader_train)\n    warmup = epochs_warmup * nbatch  # количество warmup шагов\n    nsteps = epochs * nbatch        # общее количество шагов\n\n    scheduler = CosineLRScheduler(optimizer,\n                  warmup_t=warmup, warmup_lr_init=0.0, warmup_prefix=True, # 1 эпоха warmup\n                  t_initial=(nsteps - warmup), lr_min=1e-6)                # 3 эпохи cosine\n    \n    time_val = 0.0\n    lrs = []\n\n    tb = time.time()\n    print('Epoch   loss          score   lr')\n    for iepoch in range(epochs):\n        loss_sum = 0.0\n        n_sum = 0\n\n        # Train\n        for ibatch, (img, y) in enumerate(loader_train):\n            n = y.size(0)\n            img = img.to(device)\n            y = y.to(device)\n\n            optimizer.zero_grad()\n\n            y_pred = model(img)\n            loss = criterion(y_pred.view(-1), y)\n\n            loss_train = loss.item()\n            loss_sum += n * loss_train\n            n_sum += n\n\n            loss.backward()\n\n            grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(),\n                                                       max_grad_norm)\n            optimizer.step()\n            \n            scheduler.step(iepoch * nbatch + ibatch + 1)\n            lrs.append(optimizer.param_groups[0]['lr'])            \n\n        # Оценка\n        val = evaluate(model, loader_val)\n        time_val += val['time']\n        loss_train = loss_sum / n_sum\n        lr_now = optimizer.param_groups[0]['lr']\n        dt = (time.time() - tb) / 60\n        print('Epoch %d %.4f %.4f %.4f  %.2e  %.2f min' %\n              (iepoch + 1, loss_train, val['loss'], val['score'], lr_now, dt))\n\n    dt = time.time() - tb\n    print('Training done %.2f min total, %.2f min val' % (dt / 60, time_val / 60))\n\n    # Сохранение модели\n    ofilename = 'model%d.pytorch' % (ifold + 1)\n    torch.save(model.state_dict(), ofilename)\n    print(ofilename, 'written')","metadata":{"id":"tkWJ1eXpgYWS","outputId":"ed0dd2e0-114f-4dd5-c5a6-2d8511e78359","execution":{"iopub.status.busy":"2023-02-28T04:19:15.70224Z","iopub.execute_input":"2023-02-28T04:19:15.702543Z","iopub.status.idle":"2023-02-28T08:31:41.858982Z","shell.execute_reply.started":"2023-02-28T04:19:15.702514Z","shell.execute_reply":"2023-02-28T08:31:41.857532Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"plt.title('График Learning rate')\nplt.xlabel('номер шага')\nplt.ylabel('learning rate')\nplt.plot(lrs)\nplt.show()","metadata":{"id":"gXnX0qHogYWT","execution":{"iopub.status.busy":"2023-02-28T08:31:41.863572Z","iopub.execute_input":"2023-02-28T08:31:41.863934Z","iopub.status.idle":"2023-02-28T08:31:42.152348Z","shell.execute_reply.started":"2023-02-28T08:31:41.863892Z","shell.execute_reply":"2023-02-28T08:31:42.151394Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Предсказание и сохранение в файл","metadata":{"id":"xRXxBCwRgYWU"}},{"cell_type":"code","source":"submit = pd.read_csv(di + '/sample_submission.csv')\nsubmit['target'] = 0\nfor i in range(1,6):\n    model = Model(model_name, pretrained=False)\n    filename = f'model{i}.pytorch'\n    model.to(device)\n    model.load_state_dict(torch.load(filename, map_location=device))\n    model.eval()\n\n    # Предсказание\n    dataset_test = Dataset('test', submit)\n    loader_test = torch.utils.data.DataLoader(dataset_test, batch_size=64, num_workers=num_workers, pin_memory=True)\n\n    test = evaluate(model, loader_test, compute_score=False, pbar=len(submit))\n\n    # Среднее из 5\n    submit['target'] += test['y_pred']/5\nsubmit.to_csv('submission-5folds.csv', index=False)\nprint('target range [%.2f, %.2f]' % (submit['target'].min(), submit['target'].max()))","metadata":{"id":"FMAAA0_pgYWV","execution":{"iopub.status.busy":"2023-02-28T08:31:42.15466Z","iopub.execute_input":"2023-02-28T08:31:42.155386Z","iopub.status.idle":"2023-02-28T10:42:16.777869Z","shell.execute_reply.started":"2023-02-28T08:31:42.155345Z","shell.execute_reply":"2023-02-28T10:42:16.77602Z"},"trusted":true},"execution_count":null,"outputs":[]}]}