{"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":"Experiment with Efficientnet3d","metadata":{}},{"cell_type":"code","source":"!pip3 install mlnotify","metadata":{"execution":{"iopub.status.busy":"2021-08-26T07:32:32.820742Z","iopub.execute_input":"2021-08-26T07:32:32.821062Z","iopub.status.idle":"2021-08-26T07:32:42.570022Z","shell.execute_reply.started":"2021-08-26T07:32:32.821032Z","shell.execute_reply":"2021-08-26T07:32:42.569164Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import os\nimport sys \nimport json\nimport glob\nimport random\nimport collections\nimport time\nimport re\n\nimport numpy as np\nimport pandas as pd\nimport pydicom\nfrom pydicom.pixel_data_handlers.util import apply_voi_lut\nimport cv2\nimport matplotlib.pyplot as plt\nimport seaborn as sns\n\nimport torch\nfrom torch import nn\nfrom torch.utils import data as torch_data\nfrom sklearn import model_selection as sk_model_selection\nfrom torch.nn import functional as torch_functional\nimport torch.nn.functional as F\n\nfrom sklearn.model_selection import StratifiedKFold\nfrom sklearn.metrics import roc_auc_score","metadata":{"execution":{"iopub.status.busy":"2021-08-26T07:32:42.571228Z","iopub.execute_input":"2021-08-26T07:32:42.571491Z","iopub.status.idle":"2021-08-26T07:32:45.053128Z","shell.execute_reply.started":"2021-08-26T07:32:42.571462Z","shell.execute_reply":"2021-08-26T07:32:45.052137Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"if os.path.exists(\"../input/rsna-miccai-brain-tumor-radiogenomic-classification\"):\n    data_directory = '../input/rsna-miccai-brain-tumor-radiogenomic-classification'\n    pytorch3dpath = \"../input/efficientnetpyttorch3d/EfficientNet-PyTorch-3D\"\nelse:\n    data_directory = '/media/roland/data/kaggle/rsna-miccai-brain-tumor-radiogenomic-classification'\n    pytorch3dpath = \"EfficientNet-PyTorch-3D\"\n    \nmri_types = ['FLAIR','T1w','T1wCE','T2w']\nSIZE = 256\nNUM_IMAGES = 64\n\nsys.path.append(pytorch3dpath)\nfrom efficientnet_pytorch_3d import EfficientNet3D","metadata":{"execution":{"iopub.status.busy":"2021-08-26T07:32:45.054624Z","iopub.execute_input":"2021-08-26T07:32:45.054964Z","iopub.status.idle":"2021-08-26T07:32:45.095982Z","shell.execute_reply.started":"2021-08-26T07:32:45.054932Z","shell.execute_reply":"2021-08-26T07:32:45.095207Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"img proc","metadata":{}},{"cell_type":"code","source":"\ndef load_dicom_image(path, img_size=SIZE, voi_lut=True, rotate=0):\n    dicom = pydicom.read_file(path)\n    data = dicom.pixel_array\n    if voi_lut:\n        data = apply_voi_lut(dicom.pixel_array, dicom)\n    else:\n        data = dicom.pixel_array\n        \n    if rotate > 0:\n        rot_choices = [0, cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE, cv2.ROTATE_180]\n        data = cv2.rotate(data, rot_choices[rotate])\n        \n    data = cv2.resize(data, (img_size, img_size))\n    return data\n\n\ndef load_dicom_images_3d(scan_id, num_imgs=NUM_IMAGES, img_size=SIZE, mri_type=\"FLAIR\", split=\"train\", rotate=0):\n\n    files = sorted(glob.glob(f\"{data_directory}/{split}/{scan_id}/{mri_type}/*.dcm\"), \n               key=lambda var:[int(x) if x.isdigit() else x for x in re.findall(r'[^0-9]|[0-9]+', var)])\n\n    middle = len(files)//2\n    num_imgs2 = num_imgs//2\n    p1 = max(0, middle - num_imgs2)\n    p2 = min(len(files), middle + num_imgs2)\n    img3d = np.stack([load_dicom_image(f, rotate=rotate) for f in files[p1:p2]]).T \n    if img3d.shape[-1] < num_imgs:\n        n_zero = np.zeros((img_size, img_size, num_imgs - img3d.shape[-1]))\n        img3d = np.concatenate((img3d,  n_zero), axis = -1)\n        \n    if np.min(img3d) < np.max(img3d):\n        img3d = img3d - np.min(img3d)\n        img3d = img3d / np.max(img3d)\n            \n    return np.expand_dims(img3d,0)\n\na = load_dicom_images_3d(\"00000\")\nprint(a.shape)\nprint(np.min(a), np.max(a), np.mean(a), np.median(a))\n\ndef set_seed(seed):\n    random.seed(seed)\n    os.environ[\"PYTHONHASHSEED\"] = str(seed)\n    np.random.seed(seed)\n    torch.manual_seed(seed)\n    if torch.cuda.is_available():\n        torch.cuda.manual_seed_all(seed)\n        torch.backends.cudnn.deterministic = True\n\nset_seed(12)","metadata":{"execution":{"iopub.status.busy":"2021-08-26T07:33:42.140003Z","iopub.execute_input":"2021-08-26T07:33:42.14056Z","iopub.status.idle":"2021-08-26T07:33:42.651198Z","shell.execute_reply.started":"2021-08-26T07:33:42.140513Z","shell.execute_reply":"2021-08-26T07:33:42.650176Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"train/test","metadata":{}},{"cell_type":"code","source":"train_df = pd.read_csv(f\"{data_directory}/train_labels.csv\")\ndisplay(train_df)\n\ndf_train, df_valid = sk_model_selection.train_test_split(\n    train_df, \n    test_size=0.2, \n    random_state=12, \n    stratify=train_df[\"MGMT_value\"],\n)","metadata":{"execution":{"iopub.status.busy":"2021-08-26T07:34:28.053839Z","iopub.execute_input":"2021-08-26T07:34:28.054378Z","iopub.status.idle":"2021-08-26T07:34:28.099357Z","shell.execute_reply.started":"2021-08-26T07:34:28.054317Z","shell.execute_reply":"2021-08-26T07:34:28.09839Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"model and train drivers","metadata":{}},{"cell_type":"code","source":"class Dataset(torch_data.Dataset):\n    def __init__(self, paths, targets=None, mri_type=None, label_smoothing=0.01, split=\"train\", augment=False):\n        self.paths = paths\n        self.targets = targets\n        self.mri_type = mri_type\n        self.label_smoothing = label_smoothing\n        self.split = split\n        self.augment = augment\n          \n    def __len__(self):\n        return len(self.paths)\n    \n    def __getitem__(self, index):\n        scan_id = self.paths[index]\n        if self.targets is None:\n            data = load_dicom_images_3d(str(scan_id).zfill(5), mri_type=self.mri_type[index], split=self.split)\n        else:\n            if self.augment:\n                rotation = np.random.randint(0,4)\n            else:\n                rotation = 0\n\n            data = load_dicom_images_3d(str(scan_id).zfill(5), mri_type=self.mri_type[index], split=\"train\", rotate=rotation)\n\n        if self.targets is None:\n            return {\"X\": torch.tensor(data).float(), \"id\": scan_id}\n        else:\n            y = torch.tensor(abs(self.targets[index]-self.label_smoothing), dtype=torch.float)\n            return {\"X\": torch.tensor(data).float(), \"y\": y}","metadata":{"execution":{"iopub.status.busy":"2021-08-26T07:35:22.879109Z","iopub.execute_input":"2021-08-26T07:35:22.879455Z","iopub.status.idle":"2021-08-26T07:35:22.889145Z","shell.execute_reply.started":"2021-08-26T07:35:22.879425Z","shell.execute_reply":"2021-08-26T07:35:22.887473Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class Model(nn.Module):\n    def __init__(self):\n        super().__init__()\n        self.net = EfficientNet3D.from_name(\"efficientnet-b0\", override_params={'num_classes': 2}, in_channels=1)\n        n_features = self.net._fc.in_features\n        self.net._fc = nn.Linear(in_features=n_features, out_features=1, bias=True)\n    \n    def forward(self, x):\n        out = self.net(x)\n        return out","metadata":{"execution":{"iopub.status.busy":"2021-08-26T07:35:35.581374Z","iopub.execute_input":"2021-08-26T07:35:35.581698Z","iopub.status.idle":"2021-08-26T07:35:35.587936Z","shell.execute_reply.started":"2021-08-26T07:35:35.58167Z","shell.execute_reply":"2021-08-26T07:35:35.586532Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class Trainer:\n    def __init__(\n        self, \n        model, \n        device, \n        optimizer, \n        criterion\n    ):\n        self.model = model\n        self.device = device\n        self.optimizer = optimizer\n        self.criterion = criterion\n\n        self.best_valid_score = np.inf\n        self.n_patience = 0\n        self.lastmodel = None\n        \n    def fit(self, epochs, train_loader, valid_loader, save_path, patience):        \n        for n_epoch in range(1, epochs + 1):\n            self.info_message(\"EPOCH: {}\", n_epoch)\n            \n            train_loss, train_time = self.train_epoch(train_loader)\n            valid_loss, valid_auc, valid_time = self.valid_epoch(valid_loader)\n            \n            self.info_message(\n                \"[Epoch Train: {}] loss: {:.4f}, time: {:.2f} s            \",\n                n_epoch, train_loss, train_time\n            )\n            \n            self.info_message(\n                \"[Epoch Valid: {}] loss: {:.4f}, auc: {:.4f}, time: {:.2f} s\",\n                n_epoch, valid_loss, valid_auc, valid_time\n            )\n\n            # if True:\n            #if self.best_valid_score < valid_auc: \n            if self.best_valid_score > valid_loss: \n                self.save_model(n_epoch, save_path, valid_loss, valid_auc)\n                self.info_message(\n                     \"auc improved from {:.4f} to {:.4f}. Saved model to '{}'\", \n                    self.best_valid_score, valid_loss, self.lastmodel\n                )\n                self.best_valid_score = valid_loss\n                self.n_patience = 0\n            else:\n                self.n_patience += 1\n            \n            if self.n_patience >= patience:\n                self.info_message(\"\\nValid auc didn't improve last {} epochs.\", patience)\n                break\n            \n    def train_epoch(self, train_loader):\n        self.model.train()\n        t = time.time()\n        sum_loss = 0\n\n        for step, batch in enumerate(train_loader, 1):\n            X = batch[\"X\"].to(self.device)\n            targets = batch[\"y\"].to(self.device)\n            self.optimizer.zero_grad()\n            outputs = self.model(X).squeeze(1)\n            \n            loss = self.criterion(outputs, targets)\n            loss.backward()\n\n            sum_loss += loss.detach().item()\n\n            self.optimizer.step()\n            \n            message = 'Train Step {}/{}, train_loss: {:.4f}'\n            self.info_message(message, step, len(train_loader), sum_loss/step, end=\"\\r\")\n        \n        return sum_loss/len(train_loader), int(time.time() - t)\n    \n    def valid_epoch(self, valid_loader):\n        self.model.eval()\n        t = time.time()\n        sum_loss = 0\n        y_all = []\n        outputs_all = []\n\n        for step, batch in enumerate(valid_loader, 1):\n            with torch.no_grad():\n                X = batch[\"X\"].to(self.device)\n                targets = batch[\"y\"].to(self.device)\n\n                outputs = self.model(X).squeeze(1)\n                loss = self.criterion(outputs, targets)\n\n                sum_loss += loss.detach().item()\n                y_all.extend(batch[\"y\"].tolist())\n                outputs_all.extend(torch.sigmoid(outputs).tolist())\n\n            message = 'Valid Step {}/{}, valid_loss: {:.4f}'\n            self.info_message(message, step, len(valid_loader), sum_loss/step, end=\"\\r\")\n            \n        y_all = [1 if x > 0.5 else 0 for x in y_all]\n        auc = roc_auc_score(y_all, outputs_all)\n        \n        return sum_loss/len(valid_loader), auc, int(time.time() - t)\n    \n    def save_model(self, n_epoch, save_path, loss, auc):\n        self.lastmodel = f\"{save_path}-e{n_epoch}-loss{loss:.3f}-auc{auc:.3f}.pth\"\n        torch.save(\n            {\n                \"model_state_dict\": self.model.state_dict(),\n                \"optimizer_state_dict\": self.optimizer.state_dict(),\n                \"best_valid_score\": self.best_valid_score,\n                \"n_epoch\": n_epoch,\n            },\n            self.lastmodel,\n        )\n    \n    @staticmethod\n    def info_message(message, *args, end=\"\\n\"):\n        print(message.format(*args), end=end)","metadata":{"execution":{"iopub.status.busy":"2021-08-26T07:35:54.123807Z","iopub.execute_input":"2021-08-26T07:35:54.124238Z","iopub.status.idle":"2021-08-26T07:35:54.141395Z","shell.execute_reply.started":"2021-08-26T07:35:54.124177Z","shell.execute_reply":"2021-08-26T07:35:54.14053Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Train Driver","metadata":{}},{"cell_type":"code","source":"import mlnotify\ndevice = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n\ndef train_mri_type(df_train, df_valid, mri_type):\n    if mri_type==\"all\":\n        train_list = []\n        valid_list = []\n        for mri_type in mri_types:\n            df_train.loc[:,\"MRI_Type\"] = mri_type\n            train_list.append(df_train.copy())\n            df_valid.loc[:,\"MRI_Type\"] = mri_type\n            valid_list.append(df_valid.copy())\n\n        df_train = pd.concat(train_list)\n        df_valid = pd.concat(valid_list)\n    else:\n        df_train.loc[:,\"MRI_Type\"] = mri_type\n        df_valid.loc[:,\"MRI_Type\"] = mri_type\n\n    print(df_train.shape, df_valid.shape)\n    display(df_train.head())\n    \n    train_data_retriever = Dataset(\n        df_train[\"BraTS21ID\"].values, \n        df_train[\"MGMT_value\"].values, \n        df_train[\"MRI_Type\"].values,\n        augment=True\n    )\n\n    valid_data_retriever = Dataset(\n        df_valid[\"BraTS21ID\"].values, \n        df_valid[\"MGMT_value\"].values,\n        df_valid[\"MRI_Type\"].values\n    )\n\n    train_loader = torch_data.DataLoader(\n        train_data_retriever,\n        batch_size=4,\n        shuffle=True,\n        num_workers=8,pin_memory = True\n    )\n\n    valid_loader = torch_data.DataLoader(\n        valid_data_retriever, \n        batch_size=4,\n        shuffle=False,\n        num_workers=8,pin_memory = True\n    )\n\n    model = Model()\n    model.to(device)\n\n    #checkpoint = torch.load(\"best-model-all-auc0.555.pth\")\n    #model.load_state_dict(checkpoint[\"model_state_dict\"])\n\n    #print(model)\n\n    optimizer = torch.optim.Adam(model.parameters(), lr=0.001)\n    #optimizer = torch.optim.SGD(model.parameters(), lr=0.001, momentum=0.9)\n\n    criterion = torch_functional.binary_cross_entropy_with_logits\n\n    trainer = Trainer(\n        model, \n        device, \n        optimizer, \n        criterion\n    )\n\n    history = trainer.fit(\n        10, \n        train_loader, \n        valid_loader, \n        f\"{mri_type}\", \n        10,\n    )\n    \n    return trainer.lastmodel\n\nmodelfiles = None\n\nif not modelfiles:\n    modelfiles = [train_mri_type(df_train, df_valid, m) for m in mri_types]\n    print(modelfiles)\n","metadata":{"execution":{"iopub.status.busy":"2021-08-26T07:36:37.808199Z","iopub.execute_input":"2021-08-26T07:36:37.808582Z","iopub.status.idle":"2021-08-26T07:37:31.96076Z","shell.execute_reply.started":"2021-08-26T07:36:37.80855Z","shell.execute_reply":"2021-08-26T07:37:31.957589Z"},"_kg_hide-output":true,"_kg_hide-input":true,"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"predict","metadata":{}},{"cell_type":"code","source":"def predict(modelfile, df, mri_type, split):\n    print(\"Predict:\", modelfile, mri_type, df.shape)\n    df.loc[:,\"MRI_Type\"] = mri_type\n    data_retriever = Dataset(\n        df.index.values, \n        mri_type=df[\"MRI_Type\"].values,\n        split=split\n    )\n\n    data_loader = torch_data.DataLoader(\n        data_retriever,\n        batch_size=4,\n        shuffle=False,\n        num_workers=8,\n    )\n   \n    model = Model()\n    model.to(device)\n    \n    checkpoint = torch.load(modelfile)\n    model.load_state_dict(checkpoint[\"model_state_dict\"])\n    model.eval()\n    \n    y_pred = []\n    ids = []\n\n    for e, batch in enumerate(data_loader,1):\n        print(f\"{e}/{len(data_loader)}\", end=\"\\r\")\n        with torch.no_grad():\n            tmp_pred = torch.sigmoid(model(batch[\"X\"].to(device))).cpu().numpy().squeeze()\n            if tmp_pred.size == 1:\n                y_pred.append(tmp_pred)\n            else:\n                y_pred.extend(tmp_pred.tolist())\n            ids.extend(batch[\"id\"].numpy().tolist())\n            \n    preddf = pd.DataFrame({\"BraTS21ID\": ids, \"MGMT_value\": y_pred}) \n    preddf = preddf.set_index(\"BraTS21ID\")\n    return preddf","metadata":{"execution":{"iopub.status.busy":"2021-08-26T07:38:19.285431Z","iopub.execute_input":"2021-08-26T07:38:19.285817Z","iopub.status.idle":"2021-08-26T07:38:19.295712Z","shell.execute_reply.started":"2021-08-26T07:38:19.285782Z","shell.execute_reply":"2021-08-26T07:38:19.294384Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"val","metadata":{}},{"cell_type":"code","source":"df_valid = df_valid.set_index(\"BraTS21ID\")\ndf_valid[\"MGMT_pred\"] = 0\nfor m, mtype in zip(modelfiles,  mri_types):\n    pred = predict(m, df_valid, mtype, \"train\")\n    df_valid[\"MGMT_pred\"] += pred[\"MGMT_value\"]\ndf_valid[\"MGMT_pred\"] /= len(modelfiles)\nauc = roc_auc_score(df_valid[\"MGMT_value\"], df_valid[\"MGMT_pred\"])\nprint(f\"Validation ensemble AUC: {auc:.4f}\")\nsns.displot(df_valid[\"MGMT_pred\"])","metadata":{"execution":{"iopub.status.busy":"2021-08-26T07:39:02.009417Z","iopub.execute_input":"2021-08-26T07:39:02.010814Z","iopub.status.idle":"2021-08-26T07:39:02.092811Z","shell.execute_reply.started":"2021-08-26T07:39:02.01075Z","shell.execute_reply":"2021-08-26T07:39:02.09081Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"submission","metadata":{}},{"cell_type":"code","source":"submission = pd.read_csv(f\"{data_directory}/sample_submission.csv\", index_col=\"BraTS21ID\")\n\nsubmission[\"MGMT_value\"] = 0\nfor m, mtype in zip(modelfiles, mri_types):\n    pred = predict(m, submission, mtype, split=\"test\")\n    submission[\"MGMT_value\"] += pred[\"MGMT_value\"]\n\nsubmission[\"MGMT_value\"] /= len(modelfiles)\nsubmission[\"MGMT_value\"].to_csv(\"submission.csv\")","metadata":{},"execution_count":null,"outputs":[]}]}