{"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":"!pip install segmentation_models_pytorch\n!pip install warmup_scheduler","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:17:49.804189Z","iopub.status.busy":"2023-05-05T11:17:49.803339Z","iopub.status.idle":"2023-05-05T11:18:15.288229Z","shell.execute_reply":"2023-05-05T11:18:15.286981Z"},"papermill":{"duration":25.498202,"end_time":"2023-05-05T11:18:15.291251","exception":false,"start_time":"2023-05-05T11:17:49.793049","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Imports","metadata":{"papermill":{"duration":0.013144,"end_time":"2023-05-05T11:18:15.318467","exception":false,"start_time":"2023-05-05T11:18:15.305323","status":"completed"},"tags":[]}},{"cell_type":"code","source":"import numpy as np\nimport pandas as pd\nimport scipy as sp\nfrom sklearn.metrics import roc_auc_score, accuracy_score, f1_score, log_loss\nimport matplotlib.pyplot as plt\nimport sys\nimport os\nimport gc\nimport sys\nimport pickle\nimport warnings\nimport math\nimport time\nimport random\nimport argparse\nimport importlib\nfrom tqdm.auto import tqdm\nfrom functools import partial\n\nimport torch\nimport torch.nn as nn\nfrom torch.utils.data import DataLoader, Dataset\nfrom torch.cuda.amp import autocast, GradScaler\nfrom torch.optim import Adam, SGD, AdamW\nfrom torch.optim.lr_scheduler import CosineAnnealingWarmRestarts, CosineAnnealingLR, ReduceLROnPlateau\nimport segmentation_models_pytorch as smp\nfrom warmup_scheduler import GradualWarmupScheduler\nimport cv2\n\nimport albumentations as A\nfrom albumentations.pytorch import ToTensorV2\nfrom albumentations import ImageOnlyTransform\n\nimport shutil\nfrom pathlib import Path\nfrom contextlib import contextmanager\nfrom collections import defaultdict, Counter\nimport datetime","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:15.347134Z","iopub.status.busy":"2023-05-05T11:18:15.346355Z","iopub.status.idle":"2023-05-05T11:18:21.995716Z","shell.execute_reply":"2023-05-05T11:18:21.994432Z"},"papermill":{"duration":6.666828,"end_time":"2023-05-05T11:18:21.998534","exception":false,"start_time":"2023-05-05T11:18:15.331706","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import ssl\nssl._create_default_https_context = ssl._create_unverified_context","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.018725Z","iopub.status.busy":"2023-05-05T11:18:22.017748Z","iopub.status.idle":"2023-05-05T11:18:22.023400Z","shell.execute_reply":"2023-05-05T11:18:22.022525Z"},"papermill":{"duration":0.01787,"end_time":"2023-05-05T11:18:22.025589","exception":false,"start_time":"2023-05-05T11:18:22.007719","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# sys.path.append('/kaggle/input/pretrainedmodels/pretrainedmodels-0.7.4')\n# sys.path.append('/kaggle/input/efficientnet-pytorch/EfficientNet-PyTorch-master')\n# sys.path.append('/kaggle/input/timm-pytorch-image-models/pytorch-image-models-master')\n# sys.path.append('/kaggle/input/segmentation-models-pytorch/segmentation_models.pytorch-master')","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.045068Z","iopub.status.busy":"2023-05-05T11:18:22.044788Z","iopub.status.idle":"2023-05-05T11:18:22.049065Z","shell.execute_reply":"2023-05-05T11:18:22.047952Z"},"papermill":{"duration":0.016181,"end_time":"2023-05-05T11:18:22.051331","exception":false,"start_time":"2023-05-05T11:18:22.035150","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Main Configuration Class","metadata":{"papermill":{"duration":0.008734,"end_time":"2023-05-05T11:18:22.069062","exception":false,"start_time":"2023-05-05T11:18:22.060328","status":"completed"},"tags":[]}},{"cell_type":"code","source":"class CFG:\n    # ============== comp exp name =============\n    comp_name = 'vesuvius'\n    exp_name = 'mitb2'\n    comp_dir_path = '/kaggle/input/'\n    comp_folder_name = 'vesuvius-challenge-ink-detection'\n    comp_dataset_path = f'{comp_dir_path}{comp_folder_name}/'\n\n    # ============== pred target =============\n    target_size = 1\n\n    # ============== model cfg =============\n    model_name = 'Unet'\n    backbone = 'mit_b2' #'se_resnext50_32x4d'\n\n    in_chans = 3 # 65\n    # ============== training cfg =============\n    size = 224\n    tile_size = 224\n    stride = tile_size // 4\n\n    train_batch_size = 32 # 32\n    valid_batch_size = train_batch_size * 1\n    use_amp = True\n\n    scheduler = 'GradualWarmupSchedulerV2' # 'GradualWarmupSchedulerV2' # 'CosineAnnealingLR'\n    epochs = 15 # 30\n\n    # adamW warmupあり\n    warmup_factor = 10\n    # lr = 1e-4 / warmup_factor\n    lr = 1e-4 / warmup_factor\n\n    # ============== fold =============\n    valid_id = [1, 2, 3]\n\n    # objective_cv = 'binary'  # 'binary', 'multiclass', 'regression'\n    metric_direction = 'maximize'  # maximize, 'minimize'\n    # metrics = 'dice_coef'\n\n    # ============== fixed =============\n    pretrained = True\n    inf_weight = 'best'  # 'best'\n\n    min_lr = 1e-6\n    weight_decay = 1e-6\n    max_grad_norm = 1000\n\n    print_freq = 50\n    num_workers = 2\n\n    seed = 310\n\n    # ============== set dataset path =============\n    print('set dataset path')\n\n    outputs_path = f'/kaggle/working/outputs/{comp_name}/{exp_name}/'\n\n    submission_dir = outputs_path + 'submissions/'\n    submission_path = submission_dir + f'submission_{exp_name}.csv'\n\n    model_dir = outputs_path + \\\n        f'{comp_name}-models/'\n\n    figures_dir = outputs_path + 'figures/'\n\n    log_dir = outputs_path + 'logs/'\n    log_path = log_dir + f'{exp_name}.txt'\n\n    # ============== augmentation =============\n    train_aug_list = [\n        A.Resize(size, size),\n        A.HorizontalFlip(p=0.5),\n        A.VerticalFlip(p=0.5),\n        A.RandomBrightnessContrast(p=0.75),\n        A.ShiftScaleRotate(p=0.75),\n        A.OneOf([\n                A.GaussNoise(var_limit=[10, 50]),\n                A.GaussianBlur(),\n                A.MotionBlur(),\n                ], p=0.4),\n        A.GridDistortion(num_steps=5, distort_limit=0.3, p=0.5),\n        A.CoarseDropout(max_holes=1, max_width=int(size * 0.3), max_height=int(size * 0.3), \n                        mask_fill_value=0, p=0.5),\n        A.Normalize(\n            mean= [0] * in_chans,\n            std= [1] * in_chans\n        ),\n        ToTensorV2(transpose_mask=True),\n    ]\n\n    valid_aug_list = [\n        A.Resize(size, size),\n        A.Normalize(\n            mean= [0] * in_chans,\n            std= [1] * in_chans\n        ),\n        ToTensorV2(transpose_mask=True),\n    ]\n","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.088188Z","iopub.status.busy":"2023-05-05T11:18:22.087920Z","iopub.status.idle":"2023-05-05T11:18:22.100752Z","shell.execute_reply":"2023-05-05T11:18:22.099659Z"},"papermill":{"duration":0.026776,"end_time":"2023-05-05T11:18:22.104551","exception":false,"start_time":"2023-05-05T11:18:22.077775","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Helper Functions (as usual)","metadata":{"papermill":{"duration":0.008569,"end_time":"2023-05-05T11:18:22.121837","exception":false,"start_time":"2023-05-05T11:18:22.113268","status":"completed"},"tags":[]}},{"cell_type":"code","source":"class AverageMeter(object):\n    \"\"\"Computes and stores the average and current value\"\"\"\n\n    def __init__(self):\n        self.reset()\n\n    def reset(self):\n        self.val = 0\n        self.avg = 0\n        self.sum = 0\n        self.count = 0\n\n    def update(self, val, n=1):\n        self.val = val\n        self.sum += val * n\n        self.count += n\n        self.avg = self.sum / self.count","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.140597Z","iopub.status.busy":"2023-05-05T11:18:22.140291Z","iopub.status.idle":"2023-05-05T11:18:22.146124Z","shell.execute_reply":"2023-05-05T11:18:22.145125Z"},"papermill":{"duration":0.017673,"end_time":"2023-05-05T11:18:22.148255","exception":false,"start_time":"2023-05-05T11:18:22.130582","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def init_logger(log_file):\n    from logging import getLogger, INFO, FileHandler, Formatter, StreamHandler\n    logger = getLogger(__name__)\n    logger.setLevel(INFO)\n    handler1 = StreamHandler()\n    handler1.setFormatter(Formatter(\"%(message)s\"))\n    handler2 = FileHandler(filename=log_file)\n    handler2.setFormatter(Formatter(\"%(message)s\"))\n    logger.addHandler(handler1)\n    logger.addHandler(handler2)\n    return logger\n\ndef set_seed(seed=None, cudnn_deterministic=True):\n    if seed is None:\n        seed = 310\n    \n    os.environ['PYTHONHASHSEED'] = str(seed)\n    np.random.seed(seed)\n    random.seed(seed)\n    torch.manual_seed(seed)\n    torch.cuda.manual_seed(seed)\n    torch.backends.cudnn.deterministic = cudnn_deterministic\n    torch.backends.cudnn.benchmark = False","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.166878Z","iopub.status.busy":"2023-05-05T11:18:22.166618Z","iopub.status.idle":"2023-05-05T11:18:22.173432Z","shell.execute_reply":"2023-05-05T11:18:22.172446Z"},"papermill":{"duration":0.018579,"end_time":"2023-05-05T11:18:22.175538","exception":false,"start_time":"2023-05-05T11:18:22.156959","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def make_dirs(cfg):\n    for dir in [cfg.model_dir, cfg.figures_dir, cfg.submission_dir, cfg.log_dir]:\n        os.makedirs(dir, exist_ok=True)","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.194505Z","iopub.status.busy":"2023-05-05T11:18:22.194221Z","iopub.status.idle":"2023-05-05T11:18:22.199186Z","shell.execute_reply":"2023-05-05T11:18:22.198126Z"},"papermill":{"duration":0.016631,"end_time":"2023-05-05T11:18:22.201234","exception":false,"start_time":"2023-05-05T11:18:22.184603","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def cfg_init(cfg, mode='train'):\n    set_seed(cfg.seed)\n\n    if mode == 'train':\n        make_dirs(cfg)","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.220803Z","iopub.status.busy":"2023-05-05T11:18:22.220539Z","iopub.status.idle":"2023-05-05T11:18:22.226292Z","shell.execute_reply":"2023-05-05T11:18:22.225370Z"},"papermill":{"duration":0.018242,"end_time":"2023-05-05T11:18:22.228243","exception":false,"start_time":"2023-05-05T11:18:22.210001","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"cfg_init(CFG)\n\nLogger = init_logger(log_file=CFG.log_path)\n\ndevice = torch.device('cuda' if torch.cuda.is_available() else 'cpu')","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.247633Z","iopub.status.busy":"2023-05-05T11:18:22.246800Z","iopub.status.idle":"2023-05-05T11:18:22.309326Z","shell.execute_reply":"2023-05-05T11:18:22.308313Z"},"papermill":{"duration":0.074764,"end_time":"2023-05-05T11:18:22.311844","exception":false,"start_time":"2023-05-05T11:18:22.237080","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Image and Mask (by mask, it means the inklabels.png)\n\n### The mask will act as the label for segmentation.","metadata":{"papermill":{"duration":0.008969,"end_time":"2023-05-05T11:18:22.329936","exception":false,"start_time":"2023-05-05T11:18:22.320967","status":"completed"},"tags":[]}},{"cell_type":"code","source":"def read_image_mask(fragment_id):\n\n    images = []\n    \n    mid = 65 // 2\n    start = 29\n    end = 32\n    idxs = range(start, end)\n\n    for i in tqdm(idxs):\n        \n        image = cv2.imread(CFG.comp_dataset_path + f\"train/{fragment_id}/surface_volume/{i:02}.tif\", 0)\n\n        pad0 = (CFG.tile_size - image.shape[0] % CFG.tile_size)\n        pad1 = (CFG.tile_size - image.shape[1] % CFG.tile_size)\n\n        image = np.pad(image, [(0, pad0), (0, pad1)], constant_values=0)\n\n        images.append(image)\n    images = np.stack(images, axis=2)\n\n    mask = cv2.imread(CFG.comp_dataset_path + f\"train/{fragment_id}/inklabels.png\", 0)\n    mask = np.pad(mask, [(0, pad0), (0, pad1)], constant_values=0)\n\n    mask = mask.astype('float32')\n    mask /= 255.0\n    \n    return images, mask","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.350415Z","iopub.status.busy":"2023-05-05T11:18:22.348853Z","iopub.status.idle":"2023-05-05T11:18:22.357739Z","shell.execute_reply":"2023-05-05T11:18:22.356724Z"},"papermill":{"duration":0.021078,"end_time":"2023-05-05T11:18:22.359884","exception":false,"start_time":"2023-05-05T11:18:22.338806","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def get_train_valid_dataset(valid_id):\n    train_images = []\n    train_masks = []\n\n    valid_images = []\n    valid_masks = []\n    valid_xyxys = []\n\n    for fragment_id in range(1, 4):\n\n        image, mask = read_image_mask(fragment_id)\n\n        x1_list = list(range(0, image.shape[1]-CFG.tile_size+1, CFG.stride))\n        y1_list = list(range(0, image.shape[0]-CFG.tile_size+1, CFG.stride))\n\n        for y1 in y1_list:\n            for x1 in x1_list:\n                y2 = y1 + CFG.tile_size\n                x2 = x1 + CFG.tile_size\n                # xyxys.append((x1, y1, x2, y2))\n        \n                if fragment_id == valid_id:\n                    valid_images.append(image[y1:y2, x1:x2])\n                    valid_masks.append(mask[y1:y2, x1:x2, None])\n\n                    valid_xyxys.append([x1, y1, x2, y2])\n                else:\n                    train_images.append(image[y1:y2, x1:x2])\n                    train_masks.append(mask[y1:y2, x1:x2, None])\n\n    return train_images, train_masks, valid_images, valid_masks, valid_xyxys","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.380703Z","iopub.status.busy":"2023-05-05T11:18:22.379214Z","iopub.status.idle":"2023-05-05T11:18:22.388067Z","shell.execute_reply":"2023-05-05T11:18:22.387073Z"},"papermill":{"duration":0.021229,"end_time":"2023-05-05T11:18:22.390301","exception":false,"start_time":"2023-05-05T11:18:22.369072","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Dataset Classes","metadata":{"papermill":{"duration":0.008841,"end_time":"2023-05-05T11:18:22.407843","exception":false,"start_time":"2023-05-05T11:18:22.399002","status":"completed"},"tags":[]}},{"cell_type":"code","source":"def get_transforms(data, cfg):\n    if data == 'train':\n        aug = A.Compose(cfg.train_aug_list)\n    elif data == 'valid':\n        aug = A.Compose(cfg.valid_aug_list)\n    return aug\n\nclass CustomDataset(Dataset):\n    def __init__(self, images, cfg, labels=None, transform=None):\n        self.images = images\n        self.cfg = cfg\n        self.labels = labels\n        self.transform = transform\n\n    def __len__(self):\n        return len(self.images)\n\n    def __getitem__(self, idx):\n        image = self.images[idx]\n        label = self.labels[idx]\n\n        if self.transform:\n            data = self.transform(image=image, mask=label)\n            image = data['image']\n            label = data['mask']\n\n        return image, label","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.428617Z","iopub.status.busy":"2023-05-05T11:18:22.427134Z","iopub.status.idle":"2023-05-05T11:18:22.434985Z","shell.execute_reply":"2023-05-05T11:18:22.434057Z"},"papermill":{"duration":0.02047,"end_time":"2023-05-05T11:18:22.437131","exception":false,"start_time":"2023-05-05T11:18:22.416661","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Visualization of the Data","metadata":{"papermill":{"duration":0.008855,"end_time":"2023-05-05T11:18:22.454738","exception":false,"start_time":"2023-05-05T11:18:22.445883","status":"completed"},"tags":[]}},{"cell_type":"code","source":"# train_images, train_masks, valid_images, valid_masks, valid_xyxys = get_train_valid_dataset(1)\n# valid_xyxys = np.stack(valid_xyxys)","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.474772Z","iopub.status.busy":"2023-05-05T11:18:22.473801Z","iopub.status.idle":"2023-05-05T11:18:22.478361Z","shell.execute_reply":"2023-05-05T11:18:22.477409Z"},"papermill":{"duration":0.016635,"end_time":"2023-05-05T11:18:22.480428","exception":false,"start_time":"2023-05-05T11:18:22.463793","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# plot_dataset = CustomDataset(\n#     train_images, CFG, labels=train_masks)\n\n# transform = CFG.train_aug_list\n# transform = A.Compose(\n#     [t for t in transform if not isinstance(t, (A.Normalize, ToTensorV2))])\n\n\n# plot_count = 0\n# for i in range(1000):\n\n#     image, mask = plot_dataset[i]\n#     data = transform(image=image, mask=mask)\n#     aug_image = data['image']\n#     aug_mask = data['mask']\n\n#     if mask.sum() == 0:\n#         continue\n\n#     fig, axes = plt.subplots(1, 4, figsize=(15, 8))\n#     axes[0].imshow(image[..., 0], cmap=\"gray\")\n#     axes[1].imshow(mask, cmap=\"gray\")\n#     axes[2].imshow(aug_image[..., 0], cmap=\"gray\")\n#     axes[3].imshow(aug_mask, cmap=\"gray\")\n    \n#     plt.savefig(CFG.figures_dir + f'aug_fold_{CFG.valid_id}_{plot_count}.png')\n\n#     plot_count += 1\n#     if plot_count == 5:\n#         break","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.500199Z","iopub.status.busy":"2023-05-05T11:18:22.498755Z","iopub.status.idle":"2023-05-05T11:18:22.503895Z","shell.execute_reply":"2023-05-05T11:18:22.502985Z"},"papermill":{"duration":0.016925,"end_time":"2023-05-05T11:18:22.505990","exception":false,"start_time":"2023-05-05T11:18:22.489065","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### Some memory cleaning (coz there's not infinite RAM)","metadata":{"papermill":{"duration":0.008629,"end_time":"2023-05-05T11:18:22.523314","exception":false,"start_time":"2023-05-05T11:18:22.514685","status":"completed"},"tags":[]}},{"cell_type":"code","source":"# del plot_dataset\n# del train_images\n# del train_masks\n# del valid_images\n# del valid_masks\n# del valid_xyxys\n# gc.collect()","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.542684Z","iopub.status.busy":"2023-05-05T11:18:22.541856Z","iopub.status.idle":"2023-05-05T11:18:22.546275Z","shell.execute_reply":"2023-05-05T11:18:22.545377Z"},"papermill":{"duration":0.016503,"end_time":"2023-05-05T11:18:22.548704","exception":false,"start_time":"2023-05-05T11:18:22.532201","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Model (finally the good stuff)","metadata":{"papermill":{"duration":0.008703,"end_time":"2023-05-05T11:18:22.566039","exception":false,"start_time":"2023-05-05T11:18:22.557336","status":"completed"},"tags":[]}},{"cell_type":"code","source":"class CustomModel(nn.Module):\n    def __init__(self, cfg, weight=None):\n        super().__init__()\n        self.cfg = cfg\n\n        self.encoder = smp.Unet(\n            encoder_name=cfg.backbone, \n            encoder_weights=weight,\n            in_channels=cfg.in_chans,\n            classes=cfg.target_size,\n            activation=None,\n        )\n\n    def forward(self, image):\n        output = self.encoder(image)\n        # output = output.squeeze(-1)\n        return output","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.584600Z","iopub.status.busy":"2023-05-05T11:18:22.584318Z","iopub.status.idle":"2023-05-05T11:18:22.590002Z","shell.execute_reply":"2023-05-05T11:18:22.589030Z"},"papermill":{"duration":0.017369,"end_time":"2023-05-05T11:18:22.592119","exception":false,"start_time":"2023-05-05T11:18:22.574750","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def build_model(cfg, weight=\"imagenet\"):\n    print('model_name', cfg.model_name)\n    print('backbone', cfg.backbone)\n\n    model = CustomModel(cfg, weight)\n\n    return model","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.611037Z","iopub.status.busy":"2023-05-05T11:18:22.610762Z","iopub.status.idle":"2023-05-05T11:18:22.615519Z","shell.execute_reply":"2023-05-05T11:18:22.614531Z"},"papermill":{"duration":0.016285,"end_time":"2023-05-05T11:18:22.617389","exception":false,"start_time":"2023-05-05T11:18:22.601104","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Scheduler","metadata":{"papermill":{"duration":0.008599,"end_time":"2023-05-05T11:18:22.634802","exception":false,"start_time":"2023-05-05T11:18:22.626203","status":"completed"},"tags":[]}},{"cell_type":"code","source":"class GradualWarmupSchedulerV2(GradualWarmupScheduler):\n    \"\"\"\n    https://www.kaggle.com/code/underwearfitting/single-fold-training-of-resnet200d-lb0-965\n    \"\"\"\n    def __init__(self, optimizer, multiplier, total_epoch, after_scheduler=None):\n        super(GradualWarmupSchedulerV2, self).__init__(\n            optimizer, multiplier, total_epoch, after_scheduler)\n\n    def get_lr(self):\n        if self.last_epoch > self.total_epoch:\n            if self.after_scheduler:\n                if not self.finished:\n                    self.after_scheduler.base_lrs = [\n                        base_lr * self.multiplier for base_lr in self.base_lrs]\n                    self.finished = True\n                return self.after_scheduler.get_lr()\n            return [base_lr * self.multiplier for base_lr in self.base_lrs]\n        if self.multiplier == 1.0:\n            return [base_lr * (float(self.last_epoch) / self.total_epoch) for base_lr in self.base_lrs]\n        else:\n            return [base_lr * ((self.multiplier - 1.) * self.last_epoch / self.total_epoch + 1.) for base_lr in self.base_lrs]\n\ndef get_scheduler(cfg, optimizer):\n    scheduler_cosine = torch.optim.lr_scheduler.CosineAnnealingLR(\n        optimizer, cfg.epochs, eta_min=1e-7)\n    scheduler = GradualWarmupSchedulerV2(\n        optimizer, multiplier=10, total_epoch=1, after_scheduler=scheduler_cosine)\n\n    return scheduler\n\ndef scheduler_step(scheduler, avg_val_loss, epoch):\n    scheduler.step()\n","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.653788Z","iopub.status.busy":"2023-05-05T11:18:22.653523Z","iopub.status.idle":"2023-05-05T11:18:22.662703Z","shell.execute_reply":"2023-05-05T11:18:22.661723Z"},"papermill":{"duration":0.020825,"end_time":"2023-05-05T11:18:22.664749","exception":false,"start_time":"2023-05-05T11:18:22.643924","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Initializing the Model, optimizer and the scheduler","metadata":{"papermill":{"duration":0.008637,"end_time":"2023-05-05T11:18:22.682100","exception":false,"start_time":"2023-05-05T11:18:22.673463","status":"completed"},"tags":[]}},{"cell_type":"code","source":"# model = build_model(CFG)\n# model.to(device)\n\n# optimizer = AdamW(model.parameters(), lr=CFG.lr)\n# scheduler = get_scheduler(CFG, optimizer)","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.700646Z","iopub.status.busy":"2023-05-05T11:18:22.700376Z","iopub.status.idle":"2023-05-05T11:18:22.704169Z","shell.execute_reply":"2023-05-05T11:18:22.703248Z"},"papermill":{"duration":0.015421,"end_time":"2023-05-05T11:18:22.706247","exception":false,"start_time":"2023-05-05T11:18:22.690826","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Losses","metadata":{"papermill":{"duration":0.008587,"end_time":"2023-05-05T11:18:22.723548","exception":false,"start_time":"2023-05-05T11:18:22.714961","status":"completed"},"tags":[]}},{"cell_type":"code","source":"alpha = 0.5\nbeta = 1 - alpha\n\nDiceLoss = smp.losses.DiceLoss(mode='binary')\nBCELoss = smp.losses.SoftBCEWithLogitsLoss()\nTverskyLoss = smp.losses.TverskyLoss(mode='binary', log_loss=False, alpha=alpha, beta=beta)","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.742587Z","iopub.status.busy":"2023-05-05T11:18:22.741819Z","iopub.status.idle":"2023-05-05T11:18:22.747133Z","shell.execute_reply":"2023-05-05T11:18:22.746153Z"},"papermill":{"duration":0.01692,"end_time":"2023-05-05T11:18:22.749150","exception":false,"start_time":"2023-05-05T11:18:22.732230","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def criterion(y_pred, y_true):\n    return 0.5 * BCELoss(y_pred, y_true) + 0.5 * DiceLoss(y_pred, y_true)\n#     return DiceLoss(y_pred, y_true)\n    # return 0.5 * BCELoss(y_pred, y_true) + 0.5 * TverskyLoss(y_pred, y_true)\n#     return BCELoss(y_pred, y_true)","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.767908Z","iopub.status.busy":"2023-05-05T11:18:22.767634Z","iopub.status.idle":"2023-05-05T11:18:22.773114Z","shell.execute_reply":"2023-05-05T11:18:22.772178Z"},"papermill":{"duration":0.017211,"end_time":"2023-05-05T11:18:22.775107","exception":false,"start_time":"2023-05-05T11:18:22.757896","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Training and Validation Function","metadata":{"papermill":{"duration":0.009209,"end_time":"2023-05-05T11:18:22.793142","exception":false,"start_time":"2023-05-05T11:18:22.783933","status":"completed"},"tags":[]}},{"cell_type":"code","source":"def train_fn(train_loader, model, criterion, optimizer, device):\n    model.train()\n\n    scaler = GradScaler(enabled=CFG.use_amp)\n    losses = AverageMeter()\n\n    for step, (images, labels) in tqdm(enumerate(train_loader), total=len(train_loader)):\n        images = images.to(device)\n        labels = labels.to(device)\n        batch_size = labels.size(0)\n\n        with autocast(CFG.use_amp):\n            y_preds = model(images)\n            loss = criterion(y_preds, labels)\n\n        losses.update(loss.item(), batch_size)\n        scaler.scale(loss).backward()\n\n        grad_norm = torch.nn.utils.clip_grad_norm_(\n            model.parameters(), CFG.max_grad_norm)\n\n        scaler.step(optimizer)\n        scaler.update()\n        optimizer.zero_grad()\n\n    return losses.avg\n\ndef valid_fn(valid_loader, model, criterion, device, valid_xyxys, valid_mask_gt):\n    mask_pred = np.zeros(valid_mask_gt.shape)\n    mask_count = np.zeros(valid_mask_gt.shape)\n\n    model.eval()\n    losses = AverageMeter()\n\n    for step, (images, labels) in tqdm(enumerate(valid_loader), total=len(valid_loader)):\n        images = images.to(device)\n        labels = labels.to(device)\n        batch_size = labels.size(0)\n\n        with torch.no_grad():\n            y_preds = model(images)\n            loss = criterion(y_preds, labels)\n        losses.update(loss.item(), batch_size)\n\n        # make whole mask\n        y_preds = torch.sigmoid(y_preds).to('cpu').numpy()\n        start_idx = step*CFG.valid_batch_size\n        end_idx = start_idx + batch_size\n        for i, (x1, y1, x2, y2) in enumerate(valid_xyxys[start_idx:end_idx]):\n            mask_pred[y1:y2, x1:x2] += y_preds[i].squeeze(0)\n            mask_count[y1:y2, x1:x2] += np.ones((CFG.tile_size, CFG.tile_size))\n\n    print(f'mask_count_min: {mask_count.min()}')\n    mask_pred /= mask_count\n    return losses.avg, mask_pred","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.811893Z","iopub.status.busy":"2023-05-05T11:18:22.811599Z","iopub.status.idle":"2023-05-05T11:18:22.823338Z","shell.execute_reply":"2023-05-05T11:18:22.822358Z"},"papermill":{"duration":0.023731,"end_time":"2023-05-05T11:18:22.825566","exception":false,"start_time":"2023-05-05T11:18:22.801835","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Metrics for the competition","metadata":{"papermill":{"duration":0.00856,"end_time":"2023-05-05T11:18:22.842736","exception":false,"start_time":"2023-05-05T11:18:22.834176","status":"completed"},"tags":[]}},{"cell_type":"code","source":"from sklearn.metrics import fbeta_score\n\ndef fbeta_numpy(targets, preds, beta=0.5, smooth=1e-5):\n    \"\"\"\n    https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/397288\n    \"\"\"\n    y_true_count = targets.sum()\n    ctp = preds[targets==1].sum()\n    cfp = preds[targets==0].sum()\n    beta_squared = beta * beta\n\n    c_precision = ctp / (ctp + cfp + smooth)\n    c_recall = ctp / (y_true_count + smooth)\n    dice = (1 + beta_squared) * (c_precision * c_recall) / (beta_squared * c_precision + c_recall + smooth)\n\n    return dice\n\ndef calc_fbeta(mask, mask_pred):\n    mask = mask.astype(int).flatten()\n    mask_pred = mask_pred.flatten()\n\n    best_th = 0\n    best_dice = 0\n    for th in np.array(range(10, 50+1, 5)) / 100:\n        \n        # dice = fbeta_score(mask, (mask_pred >= th).astype(int), beta=0.5)\n        dice = fbeta_numpy(mask, (mask_pred >= th).astype(int), beta=0.5)\n        print(f'th: {th}, fbeta: {dice}')\n\n        if dice > best_dice:\n            best_dice = dice\n            best_th = th\n    \n    Logger.info(f'best_th: {best_th}, fbeta: {best_dice}')\n    return best_dice, best_th\n\n\ndef calc_cv(mask_gt, mask_pred):\n    best_dice, best_th = calc_fbeta(mask_gt, mask_pred)\n\n    return best_dice, best_th","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.861904Z","iopub.status.busy":"2023-05-05T11:18:22.861025Z","iopub.status.idle":"2023-05-05T11:18:22.870427Z","shell.execute_reply":"2023-05-05T11:18:22.869498Z"},"papermill":{"duration":0.021131,"end_time":"2023-05-05T11:18:22.872531","exception":false,"start_time":"2023-05-05T11:18:22.851400","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Main","metadata":{"papermill":{"duration":0.008606,"end_time":"2023-05-05T11:18:22.889759","exception":false,"start_time":"2023-05-05T11:18:22.881153","status":"completed"},"tags":[]}},{"cell_type":"code","source":"for fold in CFG.valid_id:\n    \n    train_images, train_masks, valid_images, valid_masks, valid_xyxys = get_train_valid_dataset(fold)\n    valid_xyxys = np.stack(valid_xyxys)\n    \n    \n    train_dataset = CustomDataset(train_images, CFG, labels=train_masks,\n                                  transform=get_transforms(data='train', cfg=CFG))\n\n    valid_dataset = CustomDataset(valid_images, CFG, labels=valid_masks,\n                                  transform=get_transforms(data='valid', cfg=CFG))\n\n    train_loader = DataLoader(train_dataset,\n                              batch_size=CFG.train_batch_size,\n                              shuffle=True,\n                              num_workers=CFG.num_workers, pin_memory=True, drop_last=True,\n                              )\n\n    valid_loader = DataLoader(valid_dataset,\n                              batch_size=CFG.valid_batch_size,\n                              shuffle=False,\n                              num_workers=CFG.num_workers, pin_memory=True, drop_last=False)\n    \n    \n    valid_mask_gt = cv2.imread(CFG.comp_dataset_path + f\"train/{fold}/inklabels.png\", 0)\n    valid_mask_gt = valid_mask_gt / 255\n    pad0 = (CFG.tile_size - valid_mask_gt.shape[0] % CFG.tile_size)\n    pad1 = (CFG.tile_size - valid_mask_gt.shape[1] % CFG.tile_size)\n    valid_mask_gt = np.pad(valid_mask_gt, [(0, pad0), (0, pad1)], constant_values=0)\n    \n    \n    if CFG.metric_direction == 'minimize':\n        best_score = np.inf\n    elif CFG.metric_direction == 'maximize':\n        best_score = -1\n\n    best_loss = np.inf\n    \n    model = build_model(CFG)\n    model.to(device)\n\n    optimizer = AdamW(model.parameters(), lr=CFG.lr)\n    scheduler = get_scheduler(CFG, optimizer)\n\n    for epoch in range(CFG.epochs):\n\n        start_time = time.time()\n\n        # train\n        avg_loss = train_fn(train_loader, model, criterion, optimizer, device)\n\n        # eval\n        avg_val_loss, mask_pred = valid_fn(\n            valid_loader, model, criterion, device, valid_xyxys, valid_mask_gt)\n\n        scheduler_step(scheduler, avg_val_loss, epoch)\n#         scheduler.step()\n\n        best_dice, best_th = calc_cv(valid_mask_gt, mask_pred)\n\n        # score = avg_val_loss\n        score = best_dice\n\n        elapsed = time.time() - start_time\n\n        Logger.info(\n            f'Epoch {epoch+1} - avg_train_loss: {avg_loss:.4f}  avg_val_loss: {avg_val_loss:.4f}  time: {elapsed:.0f}s')\n        # Logger.info(f'Epoch {epoch+1} - avgScore: {avg_score:.4f}')\n        Logger.info(\n            f'Epoch {epoch+1} - avgScore: {score:.4f}')\n\n        if CFG.metric_direction == 'minimize':\n            update_best = score < best_score\n        elif CFG.metric_direction == 'maximize':\n            update_best = score > best_score\n\n        if update_best:\n            best_loss = avg_val_loss\n            best_score = score\n\n            Logger.info(\n                f'Epoch {epoch+1} - Save Best Score: {best_score:.4f} Model')\n            Logger.info(\n                f'Epoch {epoch+1} - Save Best Loss: {best_loss:.4f} Model')\n\n            torch.save({'model': model.state_dict()},\n                        CFG.model_dir + f'{CFG.model_name}_fold{fold}_best.pth')\n        torch.cuda.empty_cache()\n        gc.collect()\n    del model\n    gc.collect()\n    del optimizer\n    del scheduler\n    gc.collect()\n    torch.cuda.empty_cache()","metadata":{"execution":{"iopub.execute_input":"2023-05-05T11:18:22.908627Z","iopub.status.busy":"2023-05-05T11:18:22.908341Z","iopub.status.idle":"2023-05-05T14:01:07.199440Z","shell.execute_reply":"2023-05-05T14:01:07.198367Z"},"papermill":{"duration":9764.304164,"end_time":"2023-05-05T14:01:07.202601","exception":false,"start_time":"2023-05-05T11:18:22.898437","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{"papermill":{"duration":0.051419,"end_time":"2023-05-05T14:01:07.306370","exception":false,"start_time":"2023-05-05T14:01:07.254951","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]}]}