{"metadata":{"kernelspec":{"name":"python3","display_name":"Python 3","language":"python"},"language_info":{"name":"python","version":"3.11.13","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"colab":{"machine_shape":"hm","gpuType":"T4"},"accelerator":"GPU","kaggle":{"accelerator":"nvidiaTeslaT4","dataSources":[{"sourceId":20270,"databundleVersionId":1222630,"isSourceIdPinned":false,"sourceType":"competition"},{"sourceId":25383,"databundleVersionId":2684322,"isSourceIdPinned":false,"sourceType":"competition"},{"sourceId":9797223,"sourceType":"datasetVersion","datasetId":982170},{"sourceId":80230206,"sourceType":"kernelVersion"},{"sourceId":260782976,"sourceType":"kernelVersion"},{"sourceId":265828864,"sourceType":"kernelVersion"}],"dockerImageVersionId":31090,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"import sys\nsys.path.append(\"../input/tez-lib/\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-08T19:51:01.945647Z","iopub.execute_input":"2025-10-08T19:51:01.946026Z","iopub.status.idle":"2025-10-08T19:51:01.950753Z","shell.execute_reply.started":"2025-10-08T19:51:01.945997Z","shell.execute_reply":"2025-10-08T19:51:01.949980Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# import timm\n# from pprint import pprint\n# model_names = timm.list_models(pretrained=True)\n# pprint(model_names)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-08T19:51:01.951993Z","iopub.execute_input":"2025-10-08T19:51:01.952329Z","iopub.status.idle":"2025-10-08T19:51:01.966892Z","shell.execute_reply.started":"2025-10-08T19:51:01.952310Z","shell.execute_reply":"2025-10-08T19:51:01.966151Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import random\nimport numpy as np\nimport torch\nimport os\n\ndef seed_everything(seed):\n    random.seed(seed)\n    os.environ['PYTHONHASHSEED'] = str(seed)\n    np.random.seed(seed)\n    torch.manual_seed(seed)\n    torch.cuda.manual_seed(seed)\n    torch.backends.cudnn.deterministic = True\n    \nseed_everything(42)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-08T19:51:01.967602Z","iopub.execute_input":"2025-10-08T19:51:01.967813Z","iopub.status.idle":"2025-10-08T19:51:01.985628Z","shell.execute_reply.started":"2025-10-08T19:51:01.967797Z","shell.execute_reply":"2025-10-08T19:51:01.984874Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import torch.nn as nn\n\nfrom tez import Tez, TezConfig\nimport tez\nimport albumentations\nimport pandas as pd\nimport cv2\nimport numpy as np\nimport timm\nimport torch.nn as nn\nfrom sklearn import metrics\nimport torch\nfrom tez.callbacks import EarlyStopping\nfrom tqdm import tqdm","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-08T19:51:01.986917Z","iopub.execute_input":"2025-10-08T19:51:01.987131Z","iopub.status.idle":"2025-10-08T19:51:01.997023Z","shell.execute_reply.started":"2025-10-08T19:51:01.987114Z","shell.execute_reply":"2025-10-08T19:51:01.996380Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class args:\n    batch_size = 8\n    image_size = 384\n    epochs = 10\n    fold = 0","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-08T19:51:01.997684Z","iopub.execute_input":"2025-10-08T19:51:01.997859Z","iopub.status.idle":"2025-10-08T19:51:02.011841Z","shell.execute_reply.started":"2025-10-08T19:51:01.997844Z","shell.execute_reply":"2025-10-08T19:51:02.011315Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class CustomDataset:\n    def __init__(self, image_paths, dense_features, targets, augmentations):\n        self.image_paths = image_paths\n        self.dense_features = dense_features\n        self.targets = targets\n        self.augmentations = augmentations\n        \n    def __len__(self):\n        return len(self.image_paths)\n    \n    def __getitem__(self, item):\n        image = cv2.imread(self.image_paths[item])\n        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\n\n        if self.augmentations is not None:\n            augmented = self.augmentations(image=image)\n            image = augmented[\"image\"]\n        else:\n            image = cv2.resize(image, (args.image_size, args.image_size)).astype(np.float32)\n            image = image / 255.0\n            image = (image - np.array([0.485, 0.456, 0.406])) / np.array([0.229, 0.224, 0.225])\n\n        image = np.transpose(image, (2, 0, 1)).astype(np.float32)\n        image = np.ascontiguousarray(image)\n\n        features = self.dense_features[item, :] if (self.dense_features is not None and self.dense_features.size) else np.zeros((0,), dtype=np.float32)\n        targets = self.targets[item]\n\n        return {\n            \"image\": torch.tensor(image, dtype=torch.float),\n            \"features\": torch.tensor(features, dtype=torch.float),\n            \"targets\": torch.tensor(targets, dtype=torch.float),\n        }\n\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-08T19:51:02.012607Z","iopub.execute_input":"2025-10-08T19:51:02.012847Z","iopub.status.idle":"2025-10-08T19:51:02.029943Z","shell.execute_reply.started":"2025-10-08T19:51:02.012821Z","shell.execute_reply":"2025-10-08T19:51:02.029313Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import sklearn.metrics as skm\n\nclass CustomModel(nn.Module):\n    def __init__(self, pos_weight=None):\n        super().__init__()\n        # backbone returns features (no classifier head)\n        self.backbone = timm.create_model(\"resnet50\", pretrained=True, in_chans=3, num_classes=0, global_pool=\"avg\")\n        nf = getattr(self.backbone, \"num_features\", 2048)\n\n        self.dropout = nn.Dropout(0.5)\n        self.out = nn.Linear(nf, 1)\n\n        if pos_weight is not None:\n            # BCEWithLogitsLoss expects pos_weight on the positive class\n            self.loss_fn = nn.BCEWithLogitsLoss(pos_weight=pos_weight)\n        else:\n            self.loss_fn = nn.BCEWithLogitsLoss()\n\n        self.step_scheduler_after = \"epoch\"\n\n    def optimizer_scheduler(self):\n        opt = torch.optim.AdamW(self.parameters(), lr=1e-4, weight_decay=1e-6)\n        sch = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(opt, T_0=10, T_mult=1, eta_min=1e-6)\n        return opt, sch\n\n    def forward(self, image, features, targets=None):\n        feats = self.backbone(image)\n        feats = self.dropout(feats)\n        logits = self.out(feats).view(-1)\n\n        if targets is not None:\n            t = targets.view(-1).type_as(logits)\n            loss = self.loss_fn(logits, t)\n            with torch.no_grad():\n                probs = torch.sigmoid(logits).detach().cpu().numpy()\n                try:\n                    auc = float(skm.roc_auc_score(t.detach().cpu().numpy(), probs))\n                except Exception:\n                    auc = 0.5\n            return logits.unsqueeze(1), loss, {\n                \"binaryloss\": loss.detach(), # CHANGED from \"valid_binaryloss\"\n                \"valid_auc\": torch.tensor(auc, device=logits.device)\n            }\n        return logits.unsqueeze(1), 0, {}\n\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-08T19:51:02.030735Z","iopub.execute_input":"2025-10-08T19:51:02.030969Z","iopub.status.idle":"2025-10-08T19:51:02.049680Z","shell.execute_reply.started":"2025-10-08T19:51:02.030944Z","shell.execute_reply":"2025-10-08T19:51:02.049105Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"train_aug = albumentations.Compose(\n    [\n#         albumentations.Resize(args.image_size, args.image_size, p=1),\n        albumentations.LongestMaxSize(args.image_size, p=1),\n        albumentations.PadIfNeeded(args.image_size,args.image_size, p=1,border_mode=0),\n        \n       albumentations.HorizontalFlip(p=0.5),\n       albumentations.VerticalFlip(p=0.1),\n       albumentations.Rotate(limit=180, p=0.5),\n       albumentations.ShiftScaleRotate(\n                shift_limit=0.1, scale_limit=0.1, rotate_limit=45, p=0.5\n            ),\n        \n        albumentations.HueSaturationValue(\n            hue_shift_limit=0.2, sat_shift_limit=0.2, val_shift_limit=0.2, p=0.5\n        ),\n        albumentations.RandomBrightnessContrast(\n            brightness_limit=(-0.1, 0.1), contrast_limit=(-0.1, 0.1), p=0.5\n        ),\n        albumentations.Normalize(\n            mean=[0.485, 0.456, 0.406],\n            std=[0.229, 0.224, 0.225],\n            max_pixel_value=255.0,\n            p=1.0,\n        ),\n    ],\n    p=1.0,\n)\n\nvalid_aug = albumentations.Compose(\n    [\n#         albumentations.Resize(args.image_size, args.image_size, p=1),\n        albumentations.LongestMaxSize(args.image_size, p=1),\n        albumentations.PadIfNeeded(args.image_size,args.image_size, p=1,border_mode=0),\n        albumentations.Normalize(\n            mean=[0.485, 0.456, 0.406],\n            std=[0.229, 0.224, 0.225],\n            max_pixel_value=255.0,\n            p=1.0,\n        ),\n    ],\n    p=1.0,\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-08T19:51:02.117253Z","iopub.execute_input":"2025-10-08T19:51:02.117940Z","iopub.status.idle":"2025-10-08T19:51:02.132989Z","shell.execute_reply.started":"2025-10-08T19:51:02.117904Z","shell.execute_reply":"2025-10-08T19:51:02.132284Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"df = pd.read_csv(\"/kaggle/input/naresh-kumar-t-23bcs10006-k-folds/train_5folds.csv\")\ndf.head()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-08T19:51:02.134335Z","iopub.execute_input":"2025-10-08T19:51:02.134623Z","iopub.status.idle":"2025-10-08T19:51:02.229930Z","shell.execute_reply.started":"2025-10-08T19:51:02.134598Z","shell.execute_reply":"2025-10-08T19:51:02.229229Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"i=0\nprint(f'training fold: {i} start')\nargs.fold = 0\ndf_train = df[df.kfold != args.fold].reset_index(drop=True)\ndf_valid = df[df.kfold == args.fold].reset_index(drop=True)\npos = df_train.target.sum()\nneg = len(df_train) - pos\npos_weight = torch.tensor((neg / (pos + 1e-6))).float().cuda() if torch.cuda.is_available() else torch.tensor((neg / (pos + 1e-6))).float()\n\ndense_features = [\n]\ntrain_img_paths = [f\"/kaggle/input/siim-isic-melanoma-classification/jpeg/train/{x}.jpg\" for x in df_train[\"image_name\"].values]\nvalid_img_paths = [f\"/kaggle/input/siim-isic-melanoma-classification/jpeg/train/{x}.jpg\" for x in df_valid[\"image_name\"].values]\ntrain_dataset = CustomDataset(\n    image_paths=train_img_paths,\n    dense_features=df_train[dense_features].values,\n    targets=df_train.target.values,\n    augmentations=train_aug,\n)\n\nvalid_dataset = CustomDataset(\n    image_paths=valid_img_paths,\n    dense_features=df_valid[dense_features].values,\n    targets=df_valid.target.values,\n    augmentations=valid_aug,\n)\n\nmodel = CustomModel(pos_weight=pos_weight)\nmodel = Tez(model)\n\nconfig = TezConfig(\n    training_batch_size=args.batch_size,\n    validation_batch_size=2 * args.batch_size,\n    epochs=args.epochs,\n    step_scheduler_after=\"epoch\",\n    step_scheduler_metric=\"valid_binaryloss\", # <--- CHANGE TO \"valid_binaryloss\"\n    fp16=True,\n    val_strategy=\"epoch\",\n)\n\nes = EarlyStopping(\n    monitor=\"valid_binaryloss\",\n    model_path=f\"model_f{args.fold}.bin\",\n    patience=4,\n    mode=\"min\",\n    save_weights_only=True,\n)\n\nmodel.fit(\n    train_dataset,\n    valid_dataset=valid_dataset,\n    callbacks=[es],\n    config=config,\n)\nprint(f'training fold: {i} complete')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-08T19:51:02.230711Z","iopub.execute_input":"2025-10-08T19:51:02.230982Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null}]}