{"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":"markdown","source":"# References:\n* CenterNet Kernel from Ruslan Baynazarov https://www.kaggle.com/hocop1/centernet-baseline\n* CenterNet paper https://arxiv.org/pdf/1904.07850.pdf\n* CenterNet repository https://github.com/xingyizhou/CenterNet","metadata":{}},{"cell_type":"code","source":"!pip install efficientnet-pytorch","metadata":{"execution":{"iopub.status.busy":"2021-06-03T04:50:44.733103Z","iopub.execute_input":"2021-06-03T04:50:44.733507Z","iopub.status.idle":"2021-06-03T04:50:55.731148Z","shell.execute_reply.started":"2021-06-03T04:50:44.733416Z","shell.execute_reply":"2021-06-03T04:50:55.729755Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import numpy as np # linear algebra\nimport pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)\nimport cv2\nfrom tqdm import tqdm\nimport matplotlib.pyplot as plt\nimport seaborn as sns\nfrom functools import reduce\nimport os\nfrom scipy.optimize import minimize\nimport plotly.express as px\nfrom math import sin, cos\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.linear_model import LinearRegression\nfrom sklearn.metrics import mean_absolute_error, mean_squared_error\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torch.optim as optim\nfrom torch.optim import lr_scheduler\nfrom torch.utils.data import Dataset, DataLoader\nfrom torchvision import models\nfrom torchvision import transforms, utils\n\nfrom efficientnet_pytorch import EfficientNet\n\nPATH = '../input/pku-autonomous-driving/'","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","_kg_hide-input":true,"_kg_hide-output":true,"execution":{"iopub.status.busy":"2021-06-03T04:50:55.733253Z","iopub.execute_input":"2021-06-03T04:50:55.733736Z","iopub.status.idle":"2021-06-03T04:50:59.879054Z","shell.execute_reply.started":"2021-06-03T04:50:55.733686Z","shell.execute_reply":"2021-06-03T04:50:59.877923Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Define all functions","metadata":{}},{"cell_type":"code","source":"def imread(path, fast_mode=False):\n    img = cv2.imread(path)\n    if not fast_mode and img is not None and len(img.shape) == 3:\n        img = np.array(img[:, :, ::-1])\n    return img\n\ndef str2coords(s, names=['id', 'yaw', 'pitch', 'roll', 'x', 'y', 'z']):\n    '''\n    Input:\n        s: PredictionString (e.g. from train dataframe)\n        names: array of what to extract from the string\n    Output:\n        list of dicts with keys from `names`\n    '''\n    coords = []\n    for l in np.array(s.split()).reshape([-1, 7]):\n        coords.append(dict(zip(names, l.astype('float'))))\n        if 'id' in coords[-1]:\n            coords[-1]['id'] = int(coords[-1]['id'])\n    return coords\n\ndef rotate(x, angle):\n    x = x + angle\n    x = x - (x + np.pi) // (2 * np.pi) * 2 * np.pi\n    return x\n\ndef get_img_coords(s):\n    '''\n    Input is a PredictionString (e.g. from train dataframe)\n    Output is two arrays:\n        xs: x coordinates in the image (row)\n        ys: y coordinates in the image (column)\n    '''\n    coords = str2coords(s)\n    xs = [c['x'] for c in coords]\n    ys = [c['y'] for c in coords]\n    zs = [c['z'] for c in coords]\n    P = np.array(list(zip(xs, ys, zs))).T\n    img_p = np.dot(camera_matrix, P).T\n    img_p[:, 0] /= img_p[:, 2]\n    img_p[:, 1] /= img_p[:, 2]\n    img_xs = img_p[:, 0]\n    img_ys = img_p[:, 1]\n    img_zs = img_p[:, 2] # z = Distance from the camera\n    return img_xs, img_ys\n\ndef _regr_preprocess(regr_dict, flip=False):\n    if flip:\n        for k in ['x', 'pitch', 'roll']:\n            regr_dict[k] = -regr_dict[k]\n    for name in ['x', 'y', 'z']:\n        regr_dict[name] = regr_dict[name] / 100\n    regr_dict['roll'] = rotate(regr_dict['roll'], np.pi)\n    regr_dict['pitch_sin'] = sin(regr_dict['pitch'])\n    regr_dict['pitch_cos'] = cos(regr_dict['pitch'])\n    regr_dict.pop('pitch')\n    regr_dict.pop('id')\n    return regr_dict\n\ndef _regr_back(regr_dict):\n    for name in ['x', 'y', 'z']:\n        regr_dict[name] = regr_dict[name] * 100\n    regr_dict['roll'] = rotate(regr_dict['roll'], -np.pi)\n    \n    pitch_sin = regr_dict['pitch_sin'] / np.sqrt(regr_dict['pitch_sin']**2 + regr_dict['pitch_cos']**2)\n    pitch_cos = regr_dict['pitch_cos'] / np.sqrt(regr_dict['pitch_sin']**2 + regr_dict['pitch_cos']**2)\n    regr_dict['pitch'] = np.arccos(pitch_cos) * np.sign(pitch_sin)\n    return regr_dict\nimg = imread(PATH + 'train_images/ID_8a6e65317' + '.jpg')\nIMG_SHAPE = img.shape\nIMG_WIDTH = 1024\nIMG_HEIGHT = IMG_WIDTH // 16 * 5\nMODEL_SCALE = 8\nDISTANCE_THRESH_CLEAR = 2\n\ndef preprocess_image(img, flip=False):\n    img = img[img.shape[0] // 2:]\n    bg = np.ones_like(img) * img.mean(1, keepdims=True).astype(img.dtype)\n    bg = bg[:, :img.shape[1] // 6]\n    img = np.concatenate([bg, img, bg], 1)\n    img = cv2.resize(img, (IMG_WIDTH, IMG_HEIGHT))\n    if flip:\n        img = img[:,::-1]\n    return (img / 255).astype('float32')\n\ndef get_mask_and_regr(img, labels, flip=False):\n    mask = np.zeros([IMG_HEIGHT // MODEL_SCALE, IMG_WIDTH // MODEL_SCALE], dtype='float32')\n    regr_names = ['x', 'y', 'z', 'yaw', 'pitch', 'roll']\n    regr = np.zeros([IMG_HEIGHT // MODEL_SCALE, IMG_WIDTH // MODEL_SCALE, 7], dtype='float32')\n    coords = str2coords(labels)\n    xs, ys = get_img_coords(labels)\n    for x, y, regr_dict in zip(xs, ys, coords):\n        x, y = y, x\n        x = (x - img.shape[0] // 2) * IMG_HEIGHT / (img.shape[0] // 2) / MODEL_SCALE\n        x = np.round(x).astype('int')\n        y = (y + img.shape[1] // 6) * IMG_WIDTH / (img.shape[1] * 4/3) / MODEL_SCALE\n        y = np.round(y).astype('int')\n        if x >= 0 and x < IMG_HEIGHT // MODEL_SCALE and y >= 0 and y < IMG_WIDTH // MODEL_SCALE:\n            mask[x, y] = 1\n            regr_dict = _regr_preprocess(regr_dict, flip)\n            regr[x, y] = [regr_dict[n] for n in sorted(regr_dict)]\n    if flip:\n        mask = np.array(mask[:,::-1])\n        regr = np.array(regr[:,::-1])\n    return mask, regr\n\ndef convert_3d_to_2d(x, y, z, fx = 2304.5479, fy = 2305.8757, cx = 1686.2379, cy = 1354.9849):\n    # stolen from https://www.kaggle.com/theshockwaverider/eda-visualization-baseline\n    return x * fx / z + cx, y * fy / z + cy\n\ndef optimize_xy(r, c, x0, y0, z0, flipped=False):\n    def distance_fn(xyz):\n        x, y, z = xyz\n        xx = -x if flipped else x\n        slope_err = (xzy_slope.predict([[xx,z]])[0] - y)**2\n        x, y = convert_3d_to_2d(x, y, z)\n        y, x = x, y\n        x = (x - IMG_SHAPE[0] // 2) * IMG_HEIGHT / (IMG_SHAPE[0] // 2) / MODEL_SCALE\n        y = (y + IMG_SHAPE[1] // 6) * IMG_WIDTH / (IMG_SHAPE[1] * 4 / 3) / MODEL_SCALE\n        return max(0.2, (x-r)**2 + (y-c)**2) + max(0.4, slope_err)\n    \n    res = minimize(distance_fn, [x0, y0, z0], method='Powell')\n    x_new, y_new, z_new = res.x\n    return x_new, y_new, z_new\n\ndef clear_duplicates(coords):\n    for c1 in coords:\n        xyz1 = np.array([c1['x'], c1['y'], c1['z']])\n        for c2 in coords:\n            xyz2 = np.array([c2['x'], c2['y'], c2['z']])\n            distance = np.sqrt(((xyz1 - xyz2)**2).sum())\n            if distance < DISTANCE_THRESH_CLEAR:\n                if c1['confidence'] < c2['confidence']:\n                    c1['confidence'] = -1\n    return [c for c in coords if c['confidence'] > 0]\n\ndef extract_coords(prediction, flipped=False):\n    logits = prediction[0]\n    regr_output = prediction[1:]\n    points = np.argwhere(logits > 0)\n    col_names = sorted(['x', 'y', 'z', 'yaw', 'pitch_sin', 'pitch_cos', 'roll'])\n    coords = []\n    for r, c in points:\n        regr_dict = dict(zip(col_names, regr_output[:, r, c]))\n        coords.append(_regr_back(regr_dict))\n        coords[-1]['confidence'] = 1 / (1 + np.exp(-logits[r, c]))\n        coords[-1]['x'], coords[-1]['y'], coords[-1]['z'] = \\\n                optimize_xy(r, c,\n                            coords[-1]['x'],\n                            coords[-1]['y'],\n                            coords[-1]['z'], flipped)\n    coords = clear_duplicates(coords)\n    return coords\n\ndef coords2str(coords, names=['yaw', 'pitch', 'roll', 'x', 'y', 'z', 'confidence']):\n    s = []\n    for c in coords:\n        for n in names:\n            s.append(str(c.get(n, 0)))\n    return ' '.join(s)\n\ndef get_mesh(batch_size, shape_x, shape_y):\n    mg_x, mg_y = np.meshgrid(np.linspace(0, 1, shape_y), np.linspace(0, 1, shape_x))\n    mg_x = np.tile(mg_x[None, None, :, :], [batch_size, 1, 1, 1]).astype('float32')\n    mg_y = np.tile(mg_y[None, None, :, :], [batch_size, 1, 1, 1]).astype('float32')\n    mesh = torch.cat([torch.tensor(mg_x).to(device), torch.tensor(mg_y).to(device)], 1)\n    return mesh\n\ndef criterion(prediction, mask, regr, size_average=True):\n    # Binary mask loss\n    pred_mask = torch.sigmoid(prediction[:, 0])\n#     mask_loss = mask * (1 - pred_mask)**2 * torch.log(pred_mask + 1e-12) + (1 - mask) * pred_mask**2 * torch.log(1 - pred_mask + 1e-12)\n    mask_loss = mask * torch.log(pred_mask + 1e-12) + (1 - mask) * torch.log(1 - pred_mask + 1e-12)\n    mask_loss = -mask_loss.mean(0).sum()\n    \n    # Regression L1 loss\n    pred_regr = prediction[:, 1:]\n    regr_loss = (torch.abs(pred_regr - regr).sum(1) * mask).sum(1).sum(1) / mask.sum(1).sum(1)\n    regr_loss = regr_loss.mean(0)\n    \n    # Sum\n    loss = mask_loss + regr_loss\n    if not size_average:\n        loss *= prediction.shape[0]\n    return loss\n\ndef train_model(epoch, history=None):\n    model.train()\n\n    for batch_idx, (img_batch, mask_batch, regr_batch) in enumerate(tqdm(train_loader)):\n        img_batch = img_batch.to(device)\n        mask_batch = mask_batch.to(device)\n        regr_batch = regr_batch.to(device)\n        \n        optimizer.zero_grad()\n        output = model(img_batch)\n        loss = criterion(output, mask_batch, regr_batch)\n        if history is not None:\n            history.loc[epoch + batch_idx / len(train_loader), 'train_loss'] = loss.data.cpu().numpy()\n        \n        loss.backward()\n        \n        optimizer.step()\n        exp_lr_scheduler.step()\n    \n    print('Train Epoch: {} \\tLR: {:.6f}\\tLoss: {:.6f}'.format(epoch, optimizer.state_dict()['param_groups'][0]['lr'], loss.data))\n\ndef evaluate_model(epoch, history=None):\n    model.eval()\n    loss = 0\n    \n    with torch.no_grad():\n        for img_batch, mask_batch, regr_batch in dev_loader:\n            img_batch = img_batch.to(device)\n            mask_batch = mask_batch.to(device)\n            regr_batch = regr_batch.to(device)\n\n            output = model(img_batch)\n\n            loss += criterion(output, mask_batch, regr_batch, size_average=False).data\n    \n    loss /= len(dev_loader.dataset)\n\n    if history is not None:\n        history.loc[epoch, 'dev_loss'] = loss.cpu().numpy()\n    \n    print('Dev loss: {:.4f}'.format(loss))","metadata":{"_kg_hide-input":true,"execution":{"iopub.status.busy":"2021-06-03T04:50:59.881239Z","iopub.execute_input":"2021-06-03T04:50:59.881572Z","iopub.status.idle":"2021-06-03T04:51:00.193335Z","shell.execute_reply.started":"2021-06-03T04:50:59.881534Z","shell.execute_reply":"2021-06-03T04:51:00.192341Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Define all classes","metadata":{}},{"cell_type":"code","source":"class CarDataset(Dataset):\n    \"\"\"Car dataset.\"\"\"\n\n    def __init__(self, dataframe, root_dir, training=True, transform=None):\n        self.df = dataframe\n        self.root_dir = root_dir\n        self.transform = transform\n        self.training = training\n\n    def __len__(self):\n        return len(self.df)\n\n    def __getitem__(self, idx):\n        if torch.is_tensor(idx):\n            idx = idx.tolist()\n        \n        # Get image name\n        idx, labels = self.df.values[idx]\n        img_name = self.root_dir.format(idx)\n        \n        # Augmentation\n        flip = False\n        if self.training:\n            flip = np.random.randint(10) == 1\n        \n        # Read image\n        img0 = imread(img_name, True)\n        img = preprocess_image(img0, flip=flip)\n        img = np.rollaxis(img, 2, 0)\n        \n        # Get mask and regression maps\n        mask, regr = get_mask_and_regr(img0, labels, flip=flip)\n        regr = np.rollaxis(regr, 2, 0)\n        \n        return [img, mask, regr]\n\nclass double_conv(nn.Module):\n    '''(conv => BN => ReLU) * 2'''\n    def __init__(self, in_ch, out_ch):\n        super(double_conv, self).__init__()\n        self.conv = nn.Sequential(\n            nn.Conv2d(in_ch, out_ch, 3, padding=1),\n            nn.BatchNorm2d(out_ch),\n            nn.ReLU(inplace=True),\n            nn.Conv2d(out_ch, out_ch, 3, padding=1),\n            nn.BatchNorm2d(out_ch),\n            nn.ReLU(inplace=True)\n        )\n\n    def forward(self, x):\n        x = self.conv(x)\n        return x\n\nclass up(nn.Module):\n    def __init__(self, in_ch, out_ch, bilinear=True):\n        super(up, self).__init__()\n\n        #  would be a nice idea if the upsampling could be learned too,\n        #  but my machine do not have enough memory to handle all those weights\n        if bilinear:\n            self.up = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)\n        else:\n            self.up = nn.ConvTranspose2d(in_ch//2, in_ch//2, 2, stride=2)\n\n        self.conv = double_conv(in_ch, out_ch)\n\n    def forward(self, x1, x2=None):\n        x1 = self.up(x1)\n        \n        # input is CHW\n        diffY = x2.size()[2] - x1.size()[2]\n        diffX = x2.size()[3] - x1.size()[3]\n\n        x1 = F.pad(x1, (diffX // 2, diffX - diffX//2,\n                        diffY // 2, diffY - diffY//2))\n        \n        # for padding issues, see \n        # https://github.com/HaiyongJiang/U-Net-Pytorch-Unstructured-Buggy/commit/0e854509c2cea854e247a9c615f175f76fbb2e3a\n        # https://github.com/xiaopeng-liao/Pytorch-UNet/commit/8ebac70e633bac59fc22bb5195e513d5832fb3bd\n        \n        if x2 is not None:\n            x = torch.cat([x2, x1], dim=1)\n        else:\n            x = x1\n        x = self.conv(x)\n        return x\n\nclass MyUNet(nn.Module):\n    '''Mixture of previous classes'''\n    def __init__(self, n_classes):\n        super(MyUNet, self).__init__()\n        self.base_model = EfficientNet.from_pretrained('efficientnet-b0')\n        \n        self.conv0 = double_conv(5, 64)\n        self.conv1 = double_conv(64, 128)\n        self.conv2 = double_conv(128, 512)\n        self.conv3 = double_conv(512, 1024)\n        \n        self.mp = nn.MaxPool2d(2)\n        \n        self.up1 = up(1282 + 1024, 512)\n        self.up2 = up(512 + 512, 256)\n        self.outc = nn.Conv2d(256, n_classes, 1)\n\n    def forward(self, x):\n        batch_size = x.shape[0]\n        mesh1 = get_mesh(batch_size, x.shape[2], x.shape[3])\n        x0 = torch.cat([x, mesh1], 1)\n        x1 = self.mp(self.conv0(x0))\n        x2 = self.mp(self.conv1(x1))\n        x3 = self.mp(self.conv2(x2))\n        x4 = self.mp(self.conv3(x3))\n        \n        x_center = x[:, :, :, IMG_WIDTH // 8: -IMG_WIDTH // 8]\n        feats = self.base_model.extract_features(x_center)\n        bg = torch.zeros([feats.shape[0], feats.shape[1], feats.shape[2], feats.shape[3] // 8]).to(device)\n        feats = torch.cat([bg, feats, bg], 3)\n        \n        # Add positional info\n        mesh2 = get_mesh(batch_size, feats.shape[2], feats.shape[3])\n        feats = torch.cat([feats, mesh2], 1)\n        \n        x = self.up1(feats, x4)\n        x = self.up2(x, x3)\n        x = self.outc(x)\n        return x","metadata":{"_kg_hide-input":true,"_kg_hide-output":true,"execution":{"iopub.status.busy":"2021-06-03T04:51:00.195247Z","iopub.execute_input":"2021-06-03T04:51:00.195645Z","iopub.status.idle":"2021-06-03T04:51:00.223781Z","shell.execute_reply.started":"2021-06-03T04:51:00.195608Z","shell.execute_reply":"2021-06-03T04:51:00.22248Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Variable initialization, data loading","metadata":{}},{"cell_type":"code","source":"\ntrain = pd.read_csv(PATH + 'train.csv')\ntest = pd.read_csv(PATH + 'sample_submission.csv')\n\n# From camera.zip\ncamera_matrix = np.array([[2304.5479, 0,  1686.2379],\n                          [0, 2305.8757, 1354.9849],\n                          [0, 0, 1]], dtype=np.float32)\ncamera_matrix_inv = np.linalg.inv(camera_matrix)\n\ntrain_images_dir = PATH + 'train_images/{}.jpg'\ntest_images_dir = PATH + 'test_images/{}.jpg'\n\ndf_train, df_dev = train_test_split(train, test_size=0.01, random_state=42)\ndf_test = test\n\n# Create dataset objects\ntrain_dataset = CarDataset(df_train, train_images_dir, training=True)\ndev_dataset = CarDataset(df_dev, train_images_dir, training=False)\ntest_dataset = CarDataset(df_test, test_images_dir, training=False)\n\nBATCH_SIZE = 4\n\n# Create data generators - they will produce batches\ntrain_loader = DataLoader(dataset=train_dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=4)\ndev_loader = DataLoader(dataset=dev_dataset, batch_size=BATCH_SIZE, shuffle=False, num_workers=0)\ntest_loader = DataLoader(dataset=test_dataset, batch_size=BATCH_SIZE, shuffle=False, num_workers=0)","metadata":{"_kg_hide-input":true,"execution":{"iopub.status.busy":"2021-06-03T04:51:00.225632Z","iopub.execute_input":"2021-06-03T04:51:00.226144Z","iopub.status.idle":"2021-06-03T04:51:00.345273Z","shell.execute_reply.started":"2021-06-03T04:51:00.226094Z","shell.execute_reply":"2021-06-03T04:51:00.344001Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Gets the GPU if there is one, otherwise the cpu\ndevice = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\nprint(device)\n\nn_epochs =1\n\nmodel = MyUNet(8).to(device)\noptimizer = optim.Adam(model.parameters(), lr=0.001)\nexp_lr_scheduler = lr_scheduler.StepLR(optimizer, step_size=max(n_epochs, 10) * len(train_loader) // 3, gamma=0.1)","metadata":{"_kg_hide-input":true,"execution":{"iopub.status.busy":"2021-06-03T04:51:20.835664Z","iopub.execute_input":"2021-06-03T04:51:20.83616Z","iopub.status.idle":"2021-06-03T04:51:23.304764Z","shell.execute_reply.started":"2021-06-03T04:51:20.836126Z","shell.execute_reply":"2021-06-03T04:51:23.303706Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Training","metadata":{}},{"cell_type":"code","source":"points_df = pd.DataFrame()\nfor col in ['x', 'y', 'z', 'yaw', 'pitch', 'roll']:\n    arr = []\n    for ps in train['PredictionString']:\n        coords = str2coords(ps)\n        arr += [c[col] for c in coords]\n    points_df[col] = arr\n\nprint('len(points_df)', len(points_df))\npoints_df.head()\n\nzy_slope = LinearRegression()\nX = points_df[['z']]\ny = points_df['y']\nzy_slope.fit(X, y)\nprint('MAE without x:', mean_absolute_error(y, zy_slope.predict(X)))\n\n# Will use this model later\nxzy_slope = LinearRegression()\nX = points_df[['x', 'z']]\ny = points_df['y']\nxzy_slope.fit(X, y)\nprint('MAE with x:', mean_absolute_error(y, xzy_slope.predict(X)))\n\nprint('\\ndy/dx = {:.3f}\\ndy/dz = {:.3f}'.format(*xzy_slope.coef_))","metadata":{"execution":{"iopub.status.busy":"2021-06-03T04:51:27.160478Z","iopub.execute_input":"2021-06-03T04:51:27.160867Z","iopub.status.idle":"2021-06-03T04:51:31.381283Z","shell.execute_reply.started":"2021-06-03T04:51:27.160831Z","shell.execute_reply":"2021-06-03T04:51:31.379825Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"%%time\nimport gc\n\nhistory = pd.DataFrame()\n\nfor epoch in range(n_epochs):\n    torch.cuda.empty_cache()\n    gc.collect()\n    train_model(epoch, history)\n    evaluate_model(epoch, history)","metadata":{"_kg_hide-input":true,"_kg_hide-output":true,"pycharm":{"is_executing":true},"execution":{"iopub.status.busy":"2021-06-03T04:51:34.392568Z","iopub.execute_input":"2021-06-03T04:51:34.393056Z","iopub.status.idle":"2021-06-03T04:52:25.579035Z","shell.execute_reply.started":"2021-06-03T04:51:34.393019Z","shell.execute_reply":"2021-06-03T04:52:25.577225Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"torch.save(model.state_dict(), './model.pth')","metadata":{"pycharm":{"is_executing":true},"execution":{"iopub.status.busy":"2021-06-03T04:51:18.989774Z","iopub.status.idle":"2021-06-03T04:51:18.990725Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"history['train_loss'].iloc[100:].plot();","metadata":{"pycharm":{"is_executing":true},"execution":{"iopub.status.busy":"2021-06-03T04:51:18.991783Z","iopub.status.idle":"2021-06-03T04:51:18.992431Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"series = history.dropna()['dev_loss']\nplt.scatter(series.index, series);","metadata":{"pycharm":{"is_executing":true},"execution":{"iopub.status.busy":"2021-06-03T04:51:18.993276Z","iopub.status.idle":"2021-06-03T04:51:18.993801Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Make submission","metadata":{}},{"cell_type":"code","source":"predictions = []\n\ntest_loader = DataLoader(dataset=test_dataset, batch_size=4, shuffle=False, num_workers=4)\n\nmodel.eval()\n\nfor img, _, _ in tqdm(test_loader):\n    with torch.no_grad():\n        output = model(img.to(device))\n    output = output.data.cpu().numpy()\n    for out in output:\n        coords = extract_coords(out)\n        s = coords2str(coords)\n        predictions.append(s)","metadata":{"pycharm":{"is_executing":true},"execution":{"iopub.status.busy":"2021-06-03T04:51:18.994646Z","iopub.status.idle":"2021-06-03T04:51:18.995071Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"test = pd.read_csv(PATH + 'sample_submission.csv')\ntest['PredictionString'] = predictions\ntest.to_csv('predictions.csv', index=False)\ntest.head()","metadata":{"pycharm":{"is_executing":true},"execution":{"iopub.status.busy":"2021-06-03T04:51:18.996396Z","iopub.status.idle":"2021-06-03T04:51:18.996828Z"},"trusted":true},"execution_count":null,"outputs":[]}]}