{"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":"import os\nimport json\nimport csv\nimport random\nimport seaborn as sns\nimport pickle\nimport cv2\nimport numpy as np\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torch.optim as optim\nimport torchvision.transforms as transforms\n\nfrom PIL import Image\nfrom torch.utils.data import Dataset, DataLoader\nfrom scipy.ndimage.measurements import label\nfrom sklearn.linear_model import LogisticRegression\nfrom sklearn.neural_network import MLPClassifier\nfrom sklearn.metrics import roc_auc_score, roc_curve\n\nimport tensorflow as tf\nimport matplotlib.pyplot as plt\nimport random","metadata":{"papermill":{"duration":2.784273,"end_time":"2021-04-26T09:25:03.011883","exception":false,"start_time":"2021-04-26T09:25:00.22761","status":"completed"},"tags":[],"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"!pip install showit\nfrom showit import image as draw_image\nfrom showit import tile as draw_tile","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Dataset class","metadata":{"papermill":{"duration":0.014426,"end_time":"2021-04-26T09:25:03.046449","exception":false,"start_time":"2021-04-26T09:25:03.032023","status":"completed"},"tags":[]}},{"cell_type":"code","source":"class RefugeDataset(Dataset):\n\n    def __init__(self, root_dir, split='train', output_size=(256,256)):\n        # Define attributes\n        self.output_size = output_size\n        self.root_dir = root_dir\n        self.split = split\n        \n        # Load data index\n        with open(os.path.join(self.root_dir, self.split, 'index.json')) as f:\n            self.index = json.load(f)\n            \n        self.images = []\n        for k in range(len(self.index)):\n            print('Loading {} image {}/{}...'.format(split, k, len(self.index)), end='\\r')\n            img_name = os.path.join(self.root_dir, self.split, 'images', self.index[str(k)]['ImgName'])\n            img = np.array(Image.open(img_name).convert('RGB'))\n            img = transforms.functional.to_tensor(img)\n            img = transforms.functional.resize(img, self.output_size, interpolation=Image.BILINEAR)\n            self.images.append(img)\n            \n        # Load ground truth for 'train' and 'val' sets\n        if split != 'test':\n            self.segs = []\n            for k in range(len(self.index)):\n                print('Loading {} segmentation {}/{}...'.format(split, k, len(self.index)), end='\\r')\n                seg_name = os.path.join(self.root_dir, self.split, 'gts', self.index[str(k)]['ImgName'].split('.')[0]+'.bmp')\n                seg = np.array(Image.open(seg_name)).copy()\n                seg = 255. - seg\n                od = (seg>=127.).astype(np.float32)\n                oc = (seg>=250.).astype(np.float32)\n                od = torch.from_numpy(od[None,:,:])\n                oc = torch.from_numpy(oc[None,:,:])\n                od = transforms.functional.resize(od, self.output_size, interpolation=Image.NEAREST)\n                oc = transforms.functional.resize(oc, self.output_size, interpolation=Image.NEAREST)\n                seg = torch.cat([od, oc], dim=0)\n                self.segs.append(seg)\n                \n        print('Succesfully loaded {} dataset.'.format(split) + ' '*50)\n            \n            \n    def __len__(self):\n        return len(self.index)\n\n    def __getitem__(self, idx):\n        # Image\n        img = self.images[idx]\n    \n        # Return only images for 'test' set\n        if self.split == 'test':\n            return img\n        \n        # Else, images and ground truth\n        else:\n            # Label\n            lab = torch.tensor(self.index[str(idx)]['Label'], dtype=torch.float32)\n\n            # Segmentation masks\n            seg = self.segs[idx]\n\n            # Fovea localization\n            f_x = self.index[str(idx)]['Fovea_X']\n            f_y = self.index[str(idx)]['Fovea_Y']\n            fov = torch.FloatTensor([f_x, f_y])\n        \n            return img, lab, seg, fov, self.index[str(idx)]['ImgName']","metadata":{"papermill":{"duration":0.034267,"end_time":"2021-04-26T09:25:03.094999","exception":false,"start_time":"2021-04-26T09:25:03.060732","status":"completed"},"tags":[],"jupyter":{"source_hidden":true},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Metrics","metadata":{"papermill":{"duration":0.012912,"end_time":"2021-04-26T09:25:03.121542","exception":false,"start_time":"2021-04-26T09:25:03.10863","status":"completed"},"tags":[]}},{"cell_type":"code","source":"EPS = 1e-7\n\ndef compute_dice_coef(input, target):\n    '''\n    Compute dice score metric.\n    '''\n    batch_size = input.shape[0]\n    return sum([dice_coef_sample(input[k,:,:], target[k,:,:]) for k in range(batch_size)])/batch_size\n\ndef dice_coef_sample(input, target):\n    iflat = input.contiguous().view(-1)\n    tflat = target.contiguous().view(-1)\n    intersection = (iflat * tflat).sum()\n    return (2. * intersection) / (iflat.sum() + tflat.sum())\n\n\ndef vertical_diameter(binary_segmentation):\n    '''\n    Get the vertical diameter from a binary segmentation.\n    The vertical diameter is defined as the \"fattest\" area of the binary_segmentation parameter.\n    '''\n\n    # get the sum of the pixels in the vertical axis\n    vertical_axis_diameter = np.sum(binary_segmentation, axis=1)\n\n    # pick the maximum value\n    diameter = np.max(vertical_axis_diameter, axis=1)\n\n    # return it\n    return diameter\n\n\n\ndef vertical_cup_to_disc_ratio(od, oc):\n    '''\n    Compute the vertical cup-to-disc ratio from a given labelling map.\n    '''\n    # compute the cup diameter\n    cup_diameter = vertical_diameter(oc)\n    # compute the disc diameter\n    disc_diameter = vertical_diameter(od)\n\n    return cup_diameter / (disc_diameter + EPS)\n\ndef compute_vCDR_error(pred_od, pred_oc, gt_od, gt_oc):\n    '''\n    Compute vCDR prediction error, along with predicted vCDR and ground truth vCDR.\n    '''\n    pred_vCDR = vertical_cup_to_disc_ratio(pred_od, pred_oc)\n    gt_vCDR = vertical_cup_to_disc_ratio(gt_od, gt_oc)\n    vCDR_err = np.mean(np.abs(gt_vCDR - pred_vCDR))\n    return vCDR_err, pred_vCDR, gt_vCDR\n\n\ndef classif_eval(classif_preds, classif_gts):\n    '''\n    Compute AUC classification score.\n    '''\n    auc = roc_auc_score(classif_gts, classif_preds)\n    return auc\n\n\ndef fov_error(pred_fov, gt_fov):\n    '''\n    Fovea localization error metric (mean root squared error).\n    '''\n    err = np.sqrt(np.sum((gt_fov-pred_fov)**2, axis=1)).mean()\n    return err","metadata":{"papermill":{"duration":0.026243,"end_time":"2021-04-26T09:25:03.160687","exception":false,"start_time":"2021-04-26T09:25:03.134444","status":"completed"},"tags":[],"jupyter":{"source_hidden":true},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Post-processing functions","metadata":{"papermill":{"duration":0.012935,"end_time":"2021-04-26T09:25:03.186948","exception":false,"start_time":"2021-04-26T09:25:03.174013","status":"completed"},"tags":[]}},{"cell_type":"code","source":"def refine_seg(pred):\n    '''\n    Only retain the biggest connected component of a segmentation map.\n    '''\n    np_pred = pred.numpy()\n        \n    largest_ccs = []\n    for i in range(np_pred.shape[0]):\n        labeled, ncomponents = label(np_pred[i,:,:])\n        bincounts = np.bincount(labeled.flat)[1:]\n        if len(bincounts) == 0:\n            largest_cc = labeled == 0\n        else:\n            largest_cc = labeled == np.argmax(bincounts)+1\n        largest_cc = torch.tensor(largest_cc, dtype=torch.float32)\n        largest_ccs.append(largest_cc)\n    largest_ccs = torch.stack(largest_ccs)\n    \n    return largest_ccs","metadata":{"papermill":{"duration":0.023335,"end_time":"2021-04-26T09:25:03.226082","exception":false,"start_time":"2021-04-26T09:25:03.202747","status":"completed"},"tags":[],"jupyter":{"source_hidden":true},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Network","metadata":{"papermill":{"duration":0.013265,"end_time":"2021-04-26T09:25:03.252404","exception":false,"start_time":"2021-04-26T09:25:03.239139","status":"completed"},"tags":[]}},{"cell_type":"code","source":"class UNet(nn.Module):\n    def __init__(self, n_channels=3, n_classes=2):\n        super(UNet, self).__init__()\n        self.n_channels = n_channels\n        self.n_classes = n_classes\n        self.epoch = 0\n\n        self.inc = DoubleConv(n_channels, 64)\n        self.down1 = Down(64, 128)\n        self.down2 = Down(128, 256)\n        self.down3 = Down(256, 512)\n        factor = 2 \n        self.down4 = Down(512, 1024 // factor)\n        self.up1 = Up(1024, 512 // factor)\n        self.up2 = Up(512, 256 // factor)\n        self.up3 = Up(256, 128 // factor)\n        self.up4 = Up(128, 64)\n        self.output_layer = OutConv(64, n_classes)\n\n    def forward(self, x):\n        x1 = self.inc(x)\n        x2 = self.down1(x1)\n        x3 = self.down2(x2)\n        x4 = self.down3(x3)\n        x5 = self.down4(x4)\n        out = self.up1(x5, x4)\n        out = self.up2(out, x3)\n        out = self.up3(out, x2)\n        out = self.up4(out, x1)\n        out = self.output_layer(out)\n        out = torch.sigmoid(out)\n        return out\n\n    \nclass DoubleConv(nn.Module):\n    \"\"\"(convolution => [BN] => ReLU) * 2\"\"\"\n\n    def __init__(self, in_channels, out_channels, mid_channels=None):\n        super().__init__()\n        if not mid_channels:\n            mid_channels = out_channels\n        self.double_conv = nn.Sequential(\n            nn.Conv2d(in_channels, mid_channels, kernel_size=3, padding=1),\n            nn.BatchNorm2d(mid_channels),\n            nn.ReLU(inplace=True),\n            nn.Conv2d(mid_channels, out_channels, kernel_size=3, padding=1),\n            nn.BatchNorm2d(out_channels),\n            nn.ReLU(inplace=True)\n        )\n\n    def forward(self, x):\n        return self.double_conv(x)\n\n\nclass Down(nn.Module):\n    \"\"\"Downscaling with maxpool then double conv\"\"\"\n\n    def __init__(self, in_channels, out_channels):\n        super().__init__()\n        self.maxpool_conv = nn.Sequential(\n            nn.MaxPool2d(2),\n            DoubleConv(in_channels, out_channels)\n        )\n\n    def forward(self, x):\n        return self.maxpool_conv(x)\n\n\nclass Up(nn.Module):\n    \"\"\"Upscaling then double conv\"\"\"\n\n    def __init__(self, in_channels, out_channels):\n        super().__init__()\n\n        # Use the normal convolutions to reduce the number of channels\n        self.up = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)\n        self.conv = DoubleConv(in_channels, out_channels, in_channels // 2)\n\n\n    def forward(self, x1, x2):\n        x1 = self.up(x1)\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        x = torch.cat([x2, x1], dim=1)\n        return self.conv(x)\n\n\nclass OutConv(nn.Module):\n    '''\n    Simple convolution.\n    '''\n    def __init__(self, in_channels, out_channels):\n        super(OutConv, self).__init__()\n        self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=1)\n\n    def forward(self, x):\n        return self.conv(x)","metadata":{"papermill":{"duration":0.038018,"end_time":"2021-04-26T09:25:03.303603","exception":false,"start_time":"2021-04-26T09:25:03.265585","status":"completed"},"tags":[],"jupyter":{"source_hidden":true},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Settings","metadata":{"papermill":{"duration":0.013662,"end_time":"2021-04-26T09:25:03.330687","exception":false,"start_time":"2021-04-26T09:25:03.317025","status":"completed"},"tags":[]}},{"cell_type":"code","source":"root_dir = '/kaggle/input/eurecom-aml-2021-challenge-2/refuge_data/refuge_data'\nlr = 1e-4\nbatch_size = 8\nnum_workers = 8\ntotal_epoch = 100","metadata":{"papermill":{"duration":0.019165,"end_time":"2021-04-26T09:25:03.363242","exception":false,"start_time":"2021-04-26T09:25:03.344077","status":"completed"},"tags":[],"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Create datasets and data loaders\nAll image files are loaded in RAM in order to speed up the pipeline. Therefore, each dataset creation should take a few minutes.","metadata":{"papermill":{"duration":0.013181,"end_time":"2021-04-26T09:25:03.389864","exception":false,"start_time":"2021-04-26T09:25:03.376683","status":"completed"},"tags":[]}},{"cell_type":"code","source":"# Datasets\ntrain_set = RefugeDataset(root_dir, \n                          split='train')\nval_set = RefugeDataset(root_dir, \n                        split='val')\ntest_set = RefugeDataset(root_dir, \n                         split='test')\n\n\n# Dataloaders\ntrain_loader = DataLoader(train_set, \n                          batch_size=batch_size, \n                          shuffle=True, \n                          num_workers=num_workers,\n                          pin_memory=True,\n                         )\nval_loader = DataLoader(val_set, \n                        batch_size=batch_size, \n                        shuffle=False, \n                        num_workers=num_workers,\n                        pin_memory=True,\n                        )\ntest_loader = DataLoader(test_set, \n                        batch_size=batch_size, \n                        shuffle=False, \n                        num_workers=num_workers,\n                        pin_memory=True)","metadata":{"papermill":{"duration":299.002965,"end_time":"2021-04-26T09:30:02.405904","exception":false,"start_time":"2021-04-26T09:25:03.402939","status":"completed"},"tags":[],"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Let's see the data \n","metadata":{}},{"cell_type":"code","source":"def ratio_gl(list):\n    n=0\n    for i in range(len(list)):\n        if list[i][1] == 1 : \n            n+=1\n    return n / len(list)\n\n\nprint('Ratio of glaucaumous images :',  ratio_gl(train_set))\n","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"**It is an imbalancer data set. Working with such a distribution will have a negative impact classifying images from under represented classes. We will use data augmentation.**","metadata":{}},{"cell_type":"markdown","source":"#### Some useful functions","metadata":{}},{"cell_type":"code","source":"def show_image(image):\n    plt.figure(figsize=(10,10))\n    #Before showing image, bgr color order transformed to rgb order\n    plt.imshow(cv2.cvtColor(image, cv2.COLOR_BGR2RGB))\n    plt.xticks([])\n    plt.yticks([])\n    plt.show()\n    \ndef rotate(sample,degrees):\n    res=[]\n    for i in range(4):\n        res.append(torch.clone(sample[i]))\n    res.append(sample[4])\n    \n    for i in range(3):\n        \n        image= Image.fromarray(sample[0][i].cpu().detach().numpy())\n        rotated = Image.Image.rotate(image, degrees)\n        res[0][i]=torch.from_numpy(np.array(rotated))\n        \n    for i in range(2):\n        image= Image.fromarray(sample[2][i].cpu().detach().numpy())\n        rotated = Image.Image.rotate(image, degrees)\n        res[2][i]=torch.from_numpy(np.array(rotated))\n    return res","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"image = cv2.imread(\"/kaggle/input/eurecom-aml-2021-challenge-2/refuge_data/refuge_data/train/images/g0001.jpg\")\nshow_image(image)\n\nprint(\"Performing image denoising, in particular, using non local means denoising algorithm, because noise is expected to be a gaussian.\")\ndst = cv2.fastNlMeansDenoisingColored(image, None, 50, 20) \ndisplay = np.hstack((image, dst))\nshow_image(display)\n\n\nprint(\"Changing image color profiles\")\nHSV = cv2.cvtColor(image,cv2.COLOR_BGR2HSV)\nshow_image(HSV)\n\nprint(\"Edge detection\")\ngray = cv2.cvtColor(image,cv2.COLOR_BGR2GRAY)\ncanny_image = cv2.Canny(gray,150, 200)\n# cv2_imshow(canny_image)\n# Erosion and Dilation\nkernel = np.ones((5,5), np.uint8)\n#Dilation\ndilate_image = cv2.dilate(canny_image, kernel, iterations=1)\n# cv2_imshow(dilate_image)\n#Erosion\n# kernel = np.ones((1,1), np.uint8)\nerode_image = cv2.erode(dilate_image,kernel, iterations=1)\n# cv2_imshow(erode_image)\nnew = np.hstack((canny_image,dilate_image,erode_image))\nshow_image(new)\n\n\ndraw_tile(train_set[1][2])\n\n","metadata":{"jupyter":{"source_hidden":true},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"l = 1\nfor i,n  in zip(['none', 'nearest', 'bicubic', 'spline36', 'gaussian'], range(3)):\n    plt.figure(figsize=(10,5))\n    plt.suptitle(\"Figure %d\" %(n+1))\n    plt.subplot(121)\n    l += 1 \n    plt.imshow(train_set[1][0][1], interpolation=i, aspect='equal', cmap='gray')\n    plt.title(\"Layer %d of the first image - Interpolation %s\" % (l, i))\n    plt.subplot(122)\n    plt.imshow(train_set[1][0][n], interpolation=i, aspect='equal')\n    plt.title(\"Layer %d of the first image - Interpolation %s\" % (l, i))\n    plt.show()","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### Let's talk about the index ","metadata":{}},{"cell_type":"code","source":"import pandas as pd\nimport matplotlib.pyplot as plt\nimport matplotlib\n\nmeat_df = pd.read_json(\"/kaggle/input/eurecom-aml-2021-challenge-2/refuge_data/refuge_data/train/index.json\")\nmeat_df","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"T**he images have a similar shape.** So we do not worry about this. ","metadata":{}},{"cell_type":"code","source":"fovea_x = []\nfovea_y =[]\nlabel = []\n\nfor i in range(400):\n    fovea_x.append(meat_df[i][1])\n    fovea_y.append(meat_df[i][2])\n    label.append(meat_df[i][5])\n    \ncolors = ['red','green']\n\nfig, ax = plt.subplots()\nplt.figure(figsize=(20,20))\nfor i in range(len(label)):\n    if label[i] == 1 : \n        ax.scatter(fovea_x[i], fovea_y[i], c =label[i],cmap= matplotlib.colors.ListedColormap('red'))\n    else: \n        ax.scatter(fovea_x[i], fovea_y[i], c =label[i],cmap=matplotlib.colors.ListedColormap('blue'))\n\n        \n\nax.legend()\nax.grid(True)\nplt.show()","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### Nothing can be conclued seeing this plot, only that it seems that there no relation between fovea and a glaucomous eye","metadata":{}},{"cell_type":"markdown","source":"# Data augmentation ","metadata":{}},{"cell_type":"markdown","source":"### Data balance","metadata":{}},{"cell_type":"code","source":"\nimport tensorflow as tf\nl = len(train_set)\n\nwhile ratio_gl(train_set) < 0.5:\n    for i in range(l):\n        if train_set[i][1] == 1:\n            train_set = (*train_set, train_set[i])\n    print('New ratio', ratio_gl(train_set))\n","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### Add rotated pictures randomly ","metadata":{}},{"cell_type":"code","source":"def rotate_data(data,nb_rotation):\n    n = len(data)\n    i  = 0\n    while i <= nb_rotation :\n        j = int(random.uniform(0,n))\n        tmp = rotate(data[j],random.uniform(-30,30))\n        data = (*data, tmp)\n        i += 1\n        \nrotate_data(train_set, 50)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### Change color of some pictures\n","metadata":{}},{"cell_type":"code","source":"j = train_set[1][0]\nt = np.array(j)\nprint(cv2.UMat(t))\nprint(np.float32(j).shape)\nprint(image.shape)\n#HtV = cv2.cvtColor(np.float32(j),cv2.COLOR_BGR2HSV)\n#show_image(HtV)\n","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def change_color_toHSV(data, nb_change):\n    n = len(data)\n    i  = 0\n    while i <= nb_change & i < n:\n        j = random.uniform(0,n)\n        tmp = rotate(data[j],random.uniform(-30,30))\n        data = (*data, tmp)","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Device, model, loss and optimizer","metadata":{"papermill":{"duration":0.394978,"end_time":"2021-04-26T09:30:03.082515","exception":false,"start_time":"2021-04-26T09:30:02.687537","status":"completed"},"tags":[]}},{"cell_type":"code","source":"# Device\ndevice = torch.device(\"cuda:0\")\n\n# Network\nmodel = UNet(n_channels=3, n_classes=2).to(device)\n\n# Loss\nseg_loss = torch.nn.BCELoss(reduction='mean')\n\n# Optimizer\noptimizer = optim.Adam(model.parameters(), lr=lr)","metadata":{"papermill":{"duration":5.452715,"end_time":"2021-04-26T09:30:09.167869","exception":false,"start_time":"2021-04-26T09:30:03.715154","status":"completed"},"tags":[],"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Train for OC/OD segmentation","metadata":{"papermill":{"duration":0.281506,"end_time":"2021-04-26T09:30:09.730551","exception":false,"start_time":"2021-04-26T09:30:09.449045","status":"completed"},"tags":[]}},{"cell_type":"code","source":"epoch = []\ntrain_loss_l = []\nval_loss_l= []\ntrain_dsc_od_l= []\nval_dsc_od_l= []\ntrain_dsc_oc_l= []\nval_dsc_oc_l= []\ntrain_vCDR_error_l= []\nval_vCDR_error_l= []\ntrain_auc_l= []\nval_auc_l= []","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Define parameters\nnb_train_batches = len(train_loader)\nnb_val_batches = len(val_loader)\nnb_iter = 0\nbest_val_auc = 0.\n\nwhile model.epoch < 20:\n    # Accumulators\n    train_vCDRs, val_vCDRs = [], []\n    train_classif_gts, val_classif_gts = [], []\n    train_loss, val_loss = 0., 0.\n    train_dsc_od, val_dsc_od = 0., 0.\n    train_dsc_oc, val_dsc_oc = 0., 0.\n    train_vCDR_error, val_vCDR_error = 0., 0.\n    \n    ############\n    # TRAINING #\n    ############\n    model.train()\n    train_data = iter(train_loader)\n    \n    \n    for k in range(nb_train_batches):\n        # Loads data\n        imgs, classif_gts, seg_gts, fov_coords, names = train_data.next()\n        imgs, classif_gts, seg_gts = imgs.to(device), classif_gts.to(device), seg_gts.to(device)\n\n        # Forward pass\n        logits = model(imgs)\n        loss = seg_loss(logits, seg_gts)\n \n        # Backward pass\n        optimizer.zero_grad()\n        loss.backward()\n        optimizer.step()\n        train_loss += loss.item() / nb_train_batches\n        \n        with torch.no_grad():\n            # Compute segmentation metric\n            pred_od = refine_seg((logits[:,0,:,:]>=0.5).type(torch.int8).cpu()).to(device)\n            pred_oc = refine_seg((logits[:,1,:,:]>=0.5).type(torch.int8).cpu()).to(device)\n            gt_od = seg_gts[:,0,:,:].type(torch.int8)\n            gt_oc = seg_gts[:,1,:,:].type(torch.int8)\n            dsc_od = compute_dice_coef(pred_od, gt_od)\n            dsc_oc = compute_dice_coef(pred_oc, gt_oc)\n            train_dsc_od += dsc_od.item()/nb_train_batches\n            train_dsc_oc += dsc_oc.item()/nb_train_batches\n\n\n            # Compute and store vCDRs\n            vCDR_error, pred_vCDR, gt_vCDR = compute_vCDR_error(pred_od.cpu().numpy(), pred_oc.cpu().numpy(), gt_od.cpu().numpy(), gt_oc.cpu().numpy())\n            train_vCDRs += pred_vCDR.tolist()\n            train_vCDR_error += vCDR_error / nb_train_batches\n            train_classif_gts += classif_gts.cpu().numpy().tolist()\n            \n        # Increase iterations\n        nb_iter += 1\n        \n        # Std out\n        print('Epoch {}, iter {}/{}, loss {:.6f}'.format(model.epoch+1, k+1, nb_train_batches, loss.item()) + ' '*20, \n              end='\\r')\n    \n    \n    # Train a logistic regression on vCDRs\n    train_vCDRs = np.array(train_vCDRs).reshape(-1,1)\n    train_classif_gts = np.array(train_classif_gts)\n    clf = LogisticRegression(random_state=0, solver='lbfgs').fit(train_vCDRs, train_classif_gts)\n    train_classif_preds = clf.predict_proba(train_vCDRs)[:,1]\n    train_auc = classif_eval(train_classif_preds, train_classif_gts)\n    \n    ##############\n    # VALIDATION #\n    ##############\n    model.eval()\n    with torch.no_grad():\n        val_data = iter(val_loader)\n        for k in range(nb_val_batches):\n            # Loads data\n            imgs, classif_gts, seg_gts, fov_coords, names = val_data.next()\n            imgs, classif_gts, seg_gts = imgs.to(device), classif_gts.to(device), seg_gts.to(device)\n\n            # Forward pass\n            logits = model(imgs)\n            val_loss += seg_loss(logits, seg_gts).item() / nb_val_batches\n\n            # Std out\n            print('Validation iter {}/{}'.format(k+1, nb_val_batches) + ' '*50, \n                  end='\\r')\n            \n            # Compute segmentation metric\n            pred_od = refine_seg((logits[:,0,:,:]>=0.5).type(torch.int8).cpu()).to(device)\n            pred_oc = refine_seg((logits[:,1,:,:]>=0.5).type(torch.int8).cpu()).to(device)\n            gt_od = seg_gts[:,0,:,:].type(torch.int8)\n            gt_oc = seg_gts[:,1,:,:].type(torch.int8)\n            dsc_od = compute_dice_coef(pred_od, gt_od)\n            dsc_oc = compute_dice_coef(pred_oc, gt_oc)\n            val_dsc_od += dsc_od.item()/nb_val_batches\n            val_dsc_oc += dsc_oc.item()/nb_val_batches\n            \n            # Compute and store vCDRs\n            vCDR_error, pred_vCDR, gt_vCDR = compute_vCDR_error(pred_od.cpu().numpy(), pred_oc.cpu().numpy(), gt_od.cpu().numpy(), gt_oc.cpu().numpy())\n            val_vCDRs += pred_vCDR.tolist()\n            val_vCDR_error += vCDR_error / nb_val_batches\n            val_classif_gts += classif_gts.cpu().numpy().tolist()\n            \n\n    # Glaucoma predictions from vCDRs\n    val_vCDRs = np.array(val_vCDRs).reshape(-1,1)\n    val_classif_gts = np.array(val_classif_gts)\n    val_classif_preds = clf.predict_proba(val_vCDRs)[:,1]\n    val_auc = classif_eval(val_classif_preds, val_classif_gts)\n        \n    # Validation results\n    print('VALIDATION epoch {}'.format(model.epoch+1)+' '*50)\n    print('LOSSES: {:.4f} (train), {:.4f} (val)'.format(train_loss, val_loss))\n    print('OD segmentation (Dice Score): {:.4f} (train), {:.4f} (val)'.format(train_dsc_od, val_dsc_od))\n    print('OC segmentation (Dice Score): {:.4f} (train), {:.4f} (val)'.format(train_dsc_oc, val_dsc_oc))\n    print('vCDR error: {:.4f} (train), {:.4f} (val)'.format(train_vCDR_error, val_vCDR_error))\n    print('Classification (AUC): {:.4f} (train), {:.4f} (val)'.format(train_auc, val_auc))\n    \n    # Save model if best validation AUC is reached\n    if val_auc > best_val_auc:\n        torch.save(model.state_dict(), '/kaggle/working/best_AUC_weights.pth')\n        with open('/kaggle/working/best_AUC_classifier.pkl', 'wb') as clf_file:\n            pickle.dump(clf, clf_file)\n        best_val_auc = val_auc\n        print('Best validation AUC reached. Saved model weights and classifier.')\n    print('_'*50)\n        \n    # End of epoch\n    epoch.append(model.epoch)\n    train_loss_l.append(train_loss)\n    val_loss_l.append(val_loss)\n    train_dsc_od_l.append(train_dsc_od)\n    val_dsc_od_l.append(val_dsc_od)\n    train_dsc_oc_l.append(train_dsc_oc)\n    val_dsc_oc_l.append(val_dsc_oc)\n    train_vCDR_error_l.append(train_vCDR_error)\n    val_vCDR_error_l.append(val_vCDR_error)\n    train_auc_l.append(train_auc)\n    val_auc_l.append(val_auc)\n    model.epoch += 1\n    \n    plt.plot(epoch,train_auc_l)\n    plt.plot(epoch,val_auc_l)\n    plt.show()\n","metadata":{"papermill":{"duration":1811.719552,"end_time":"2021-04-26T10:00:21.732656","exception":false,"start_time":"2021-04-26T09:30:10.013104","status":"completed"},"tags":[],"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Load best model + classifier","metadata":{"papermill":{"duration":2.01405,"end_time":"2021-04-26T10:00:25.741981","exception":false,"start_time":"2021-04-26T10:00:23.727931","status":"completed"},"tags":[]}},{"cell_type":"code","source":"# Load model and classifier\nmodel = UNet(n_channels=3, n_classes=2).to(device)\nmodel.load_state_dict(torch.load('/kaggle/working/best_AUC_weights.pth'))\nwith open('/kaggle/working/best_AUC_classifier.pkl', 'rb') as clf_file:\n    clf = pickle.load(clf_file)","metadata":{"papermill":{"duration":2.471641,"end_time":"2021-04-26T10:00:30.233875","exception":false,"start_time":"2021-04-26T10:00:27.762234","status":"completed"},"tags":[],"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Check performance is maintained on validation","metadata":{"papermill":{"duration":2.042081,"end_time":"2021-04-26T10:00:34.261912","exception":false,"start_time":"2021-04-26T10:00:32.219831","status":"completed"},"tags":[]}},{"cell_type":"code","source":"model.eval()\nval_vCDRs = []\nval_classif_gts = []\nval_loss = 0.\nval_dsc_od = 0.\nval_dsc_oc = 0.\nval_vCDR_error = 0.\nwith torch.no_grad():\n    val_data = iter(val_loader)\n    for k in range(nb_val_batches):\n        # Loads data\n        imgs, classif_gts, seg_gts, fov_coords, names = val_data.next()\n        imgs, classif_gts, seg_gts = imgs.to(device), classif_gts.to(device), seg_gts.to(device)\n\n        # Forward pass\n        logits = model(imgs)\n        val_loss += seg_loss(logits, seg_gts).item() / nb_val_batches\n\n        # Std out\n        print('Validation iter {}/{}'.format(k+1, nb_val_batches) + ' '*50, \n              end='\\r')\n\n        # Compute segmentation metric\n        pred_od = refine_seg((logits[:,0,:,:]>=0.5).type(torch.int8).cpu()).to(device)\n        pred_oc = refine_seg((logits[:,1,:,:]>=0.5).type(torch.int8).cpu()).to(device)\n        gt_od = seg_gts[:,0,:,:].type(torch.int8)\n        gt_oc = seg_gts[:,1,:,:].type(torch.int8)\n        dsc_od = compute_dice_coef(pred_od, gt_od)\n        dsc_oc = compute_dice_coef(pred_oc, gt_oc)\n        val_dsc_od += dsc_od.item()/nb_val_batches\n        val_dsc_oc += dsc_oc.item()/nb_val_batches\n\n        # Compute and store vCDRs\n        vCDR_error, pred_vCDR, gt_vCDR = compute_vCDR_error(pred_od.cpu().numpy(), pred_oc.cpu().numpy(), gt_od.cpu().numpy(), gt_oc.cpu().numpy())\n        val_vCDRs += pred_vCDR.tolist()\n        val_vCDR_error += vCDR_error / nb_val_batches\n        val_classif_gts += classif_gts.cpu().numpy().tolist()\n\n\n# Glaucoma predictions from vCDRs\nval_vCDRs = np.array(val_vCDRs).reshape(-1,1)\nval_classif_gts = np.array(val_classif_gts)\nval_classif_preds = clf.predict_proba(val_vCDRs)[:,1]\nval_auc = classif_eval(val_classif_preds, val_classif_gts)\n\n# Validation results\nprint('VALIDATION '+' '*50)\nprint('LOSSES: {:.4f} (val)'.format(val_loss))\nprint('OD segmentation (Dice Score): {:.4f} (val)'.format(val_dsc_od))\nprint('OC segmentation (Dice Score): {:.4f} (val)'.format(val_dsc_oc))\nprint('vCDR error: {:.4f} (val)'.format(val_vCDR_error))\nprint('Classification (AUC): {:.4f} (val)'.format(val_auc))","metadata":{"papermill":{"duration":7.689961,"end_time":"2021-04-26T10:00:43.945243","exception":false,"start_time":"2021-04-26T10:00:36.255282","status":"completed"},"tags":[],"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Predictions on test set","metadata":{"papermill":{"duration":1.995085,"end_time":"2021-04-26T10:00:47.967262","exception":false,"start_time":"2021-04-26T10:00:45.972177","status":"completed"},"tags":[]}},{"cell_type":"code","source":"nb_test_batches = len(test_loader)\nmodel.eval()\ntest_vCDRs = []\nwith torch.no_grad():\n    test_data = iter(test_loader)\n    for k in range(nb_test_batches):\n        # Loads data\n        imgs = test_data.next()\n        imgs = imgs.to(device)\n\n        # Forward pass\n        logits = model(imgs)\n\n        # Std out\n        print('Test iter {}/{}'.format(k+1, nb_test_batches) + ' '*50, \n              end='\\r')\n            \n        # Compute segmentation\n        pred_od = refine_seg((logits[:,0,:,:]>=0.5).type(torch.int8).cpu()).to(device)\n        pred_oc = refine_seg((logits[:,1,:,:]>=0.5).type(torch.int8).cpu()).to(device)\n            \n        # Compute and store vCDRs\n        pred_vCDR = vertical_cup_to_disc_ratio(pred_od.cpu().numpy(), pred_oc.cpu().numpy())\n        test_vCDRs += pred_vCDR.tolist()\n            \n\n    # Glaucoma predictions from vCDRs\n    test_vCDRs = np.array(test_vCDRs).reshape(-1,1)\n    test_classif_preds = clf.predict_proba(test_vCDRs)[:,1]\n    \n# Prepare and save .csv file\ndef create_submission_csv(prediction, submission_filename='/kaggle/working/submission.csv'):\n    \"\"\"Create a sumbission file in the appropriate format for evaluation.\n\n    :param\n    prediction: list of predictions (ex: [0.12720, 0.89289, ..., 0.29829])\n    \"\"\"\n    \n    with open(submission_filename, mode='w') as csv_file:\n        fieldnames = ['Id', 'Predicted']\n        writer = csv.DictWriter(csv_file, fieldnames=fieldnames)\n        writer.writeheader()\n\n        for i, p in enumerate(prediction):\n            writer.writerow({'Id': \"T{:04d}\".format(i+1), 'Predicted': '{:f}'.format(p)})\n\ncreate_submission_csv(test_classif_preds)\n\n# The submission.csv file is under /kaggle/working/submission.csv.\n# If you want to submit it, you should download it before closing the current kernel.","metadata":{"papermill":{"duration":7.163045,"end_time":"2021-04-26T10:00:57.140044","exception":false,"start_time":"2021-04-26T10:00:49.976999","status":"completed"},"tags":[],"trusted":true},"execution_count":null,"outputs":[]}]}