{"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":"# About this notebook\n- PyTorch TResNet starter code\n- single fold\n- 3 epochs\n\n# Improvements maybe\n- Use ArcFace or add triplet loss with cross entropy for score improvement\n\n# acknowledgement\n- Y.NAKAMA great [notebook](https://www.kaggle.com/yasufuminakama/herbarium-2020-pytorch-resnet18-train/notebook)\n- Hussam Lawen great [discussion topic](https://www.kaggle.com/c/herbarium-2020-fgvc7/discussion/154186)\n\nIf this notebook is helpful, feel free to upvote :)","metadata":{}},{"cell_type":"code","source":"!pip install -q --upgrade wandb\n!pip install -q ttach\n!pip install  timm\n!pip install git+https://github.com/mapillary/inplace_abn.git@v1.0.12","metadata":{"execution":{"iopub.status.busy":"2022-03-16T01:46:30.505188Z","iopub.execute_input":"2022-03-16T01:46:30.505844Z","iopub.status.idle":"2022-03-16T01:48:18.491979Z","shell.execute_reply.started":"2022-03-16T01:46:30.505743Z","shell.execute_reply":"2022-03-16T01:48:18.491052Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Data Loading","metadata":{}},{"cell_type":"code","source":"import os\nimport gc\nimport numpy as np\nimport pandas as pd\nfrom PIL import Image\nimport cv2 as cv\nfrom matplotlib import pyplot as plt\nimport seaborn as sns\n\npd.options.display.max_columns = 300","metadata":{"execution":{"iopub.status.busy":"2022-03-16T01:48:18.494288Z","iopub.execute_input":"2022-03-16T01:48:18.494625Z","iopub.status.idle":"2022-03-16T01:48:19.890851Z","shell.execute_reply.started":"2022-03-16T01:48:18.49457Z","shell.execute_reply":"2022-03-16T01:48:19.890032Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train = pd.read_csv('../input/herbarium-2022-pandas/train.csv')","metadata":{"execution":{"iopub.status.busy":"2022-03-16T01:48:19.892113Z","iopub.execute_input":"2022-03-16T01:48:19.892957Z","iopub.status.idle":"2022-03-16T01:48:22.289291Z","shell.execute_reply.started":"2022-03-16T01:48:19.892911Z","shell.execute_reply":"2022-03-16T01:48:22.288465Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train[\"category\"].value_counts()","metadata":{"execution":{"iopub.status.busy":"2022-03-16T01:48:22.291362Z","iopub.execute_input":"2022-03-16T01:48:22.291635Z","iopub.status.idle":"2022-03-16T01:48:22.329504Z","shell.execute_reply.started":"2022-03-16T01:48:22.291599Z","shell.execute_reply":"2022-03-16T01:48:22.328603Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Quick EDA","metadata":{}},{"cell_type":"code","source":"for i in range(5):\n    image = cv.imread(train.loc[i, 'directory'])\n    image = cv.cvtColor(image, cv.COLOR_BGR2RGB)\n    target = train.loc[i, 'category']\n    plt.imshow(image)\n    plt.title(f\"target: {target}\")\n    plt.show()","metadata":{"execution":{"iopub.status.busy":"2022-03-16T01:48:22.331183Z","iopub.execute_input":"2022-03-16T01:48:22.331469Z","iopub.status.idle":"2022-03-16T01:48:23.728618Z","shell.execute_reply.started":"2022-03-16T01:48:22.331431Z","shell.execute_reply":"2022-03-16T01:48:23.727916Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"from sklearn import preprocessing\n\nle = preprocessing.LabelEncoder()\nle.fit(train['category'])\ntrain['category'] = le.transform(train['category'])","metadata":{"execution":{"iopub.status.busy":"2022-03-16T01:48:23.729876Z","iopub.execute_input":"2022-03-16T01:48:23.730229Z","iopub.status.idle":"2022-03-16T01:48:23.879999Z","shell.execute_reply.started":"2022-03-16T01:48:23.730194Z","shell.execute_reply":"2022-03-16T01:48:23.879227Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Directory settings","metadata":{}},{"cell_type":"code","source":"# ====================================================\n# Directory settings\n# ====================================================\nimport os\n\nOUTPUT_DIR = './'\nif not os.path.exists(OUTPUT_DIR):\n    os.makedirs(OUTPUT_DIR)","metadata":{"execution":{"iopub.status.busy":"2022-03-16T01:48:23.881663Z","iopub.execute_input":"2022-03-16T01:48:23.882161Z","iopub.status.idle":"2022-03-16T01:48:23.887471Z","shell.execute_reply.started":"2022-03-16T01:48:23.882121Z","shell.execute_reply":"2022-03-16T01:48:23.88666Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Configuration","metadata":{}},{"cell_type":"code","source":"class CFG:\n    apex=False\n    debug=False\n    print_freq=100\n    size=128\n    num_workers=8\n    scheduler='CosineAnnealingLR' # ['ReduceLROnPlateau', 'CosineAnnealingLR', 'CosineAnnealingWarmRestarts','OneCycleLR']\n    epochs=3\n    # CosineAnnealingLR params\n    cosanneal_params={\n        'T_max':4,\n        'eta_min':1e-5,\n        'last_epoch':-1\n    }\n    #ReduceLROnPlateau params\n    reduce_params={\n        'mode':'min',\n        'factor':0.2,\n        'patience':4,\n        'eps':1e-6,\n        'verbose':True\n    }\n    # CosineAnnealingWarmRestarts params\n    cosanneal_res_params={\n        'T_0':3,\n        'eta_min':1e-6,\n        'T_mult':1,\n        'last_epoch':-1\n    }\n    onecycle_params={\n        'pct_start':0.1,\n        'div_factor':1e2,\n        'max_lr':1e-3\n    }\n    batch_size=32\n    lr=1e-3\n    weight_decay=1e-5\n    gradient_accumulation_steps=1\n    max_grad_norm=1000\n    target_size=train[\"category\"].shape[0]\n    nfolds=2\n    trn_folds=[0]\n    model_name='tresnet_m'     #'vit_base_patch32_224_in21k' 'tf_efficientnetv2_b0' 'resnext50_32x4d'\n    train=True\n    early_stop=True\n    target_col=\"category\"\n    fc_dim=512\n    early_stopping_steps=5\n    grad_cam=False\n    seed=42\n    \nif CFG.debug:\n    CFG.epochs=1\n    train=train.sample(n=1000, random_state=CFG.seed).reset_index(drop=True)","metadata":{"execution":{"iopub.status.busy":"2022-03-16T01:48:23.888909Z","iopub.execute_input":"2022-03-16T01:48:23.88938Z","iopub.status.idle":"2022-03-16T01:48:23.902043Z","shell.execute_reply.started":"2022-03-16T01:48:23.889342Z","shell.execute_reply":"2022-03-16T01:48:23.901037Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Library","metadata":{}},{"cell_type":"code","source":"# ====================================================\n# Library\n# ====================================================\nimport sys\nimport os\nimport math\nimport time\nimport random\nimport shutil\nfrom pathlib import Path\nfrom contextlib import contextmanager\nfrom collections import defaultdict, Counter\n\nimport scipy as sp\nimport numpy as np\nimport pandas as pd\n\nfrom sklearn import preprocessing\nfrom sklearn.metrics import roc_auc_score, roc_curve, f1_score, accuracy_score\nfrom sklearn.model_selection import StratifiedKFold, GroupKFold, KFold\n\nfrom tqdm.auto import tqdm\nfrom functools import partial\n\nimport cv2\nfrom PIL import Image\nfrom PIL import ImageFile\n# sometimes, you will have images without an ending bit\n# this takes care of those kind of (corrupt) images\nImageFile.LOAD_TRUNCATED_IMAGES = True\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom torch.optim import Adam, SGD\nfrom torch.optim.optimizer import Optimizer\nimport torchvision.models as models\nfrom torch.nn.parameter import Parameter\nfrom torch.utils.data import DataLoader, Dataset\nfrom torch.optim.lr_scheduler import CosineAnnealingWarmRestarts, CosineAnnealingLR, ReduceLROnPlateau\n\n\nimport albumentations as A\nfrom albumentations.pytorch import ToTensorV2\nfrom albumentations import ImageOnlyTransform\n\n\nimport timm\n\nfrom torch.cuda.amp import autocast, GradScaler\n\nimport warnings\nwarnings.filterwarnings('ignore')\n\ndevice = torch.device('cuda' if torch.cuda.is_available() else 'cpu')","metadata":{"execution":{"iopub.status.busy":"2022-03-16T01:48:23.903726Z","iopub.execute_input":"2022-03-16T01:48:23.904383Z","iopub.status.idle":"2022-03-16T01:48:26.165446Z","shell.execute_reply.started":"2022-03-16T01:48:23.904346Z","shell.execute_reply":"2022-03-16T01:48:26.164429Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# W&B","metadata":{}},{"cell_type":"code","source":"from kaggle_secrets import UserSecretsClient\nuser_secrets = UserSecretsClient()\nwandb_api = user_secrets.get_secret(\"wandb_key\")\n\nimport wandb\nwandb.login(key=wandb_api)\n\ndef class2dict(f):\n    return dict((name, getattr(f, name)) for name in dir(f) if not name.startswith('__'))\n\nrun = wandb.init(project=\"Herbarium 2022\", \n                 name=\"resnext50_32x4d\",\n                 config=class2dict(CFG),\n                 group=CFG.model_name,\n                 job_type=\"train\")","metadata":{"execution":{"iopub.status.busy":"2022-03-16T01:48:26.169415Z","iopub.execute_input":"2022-03-16T01:48:26.169702Z","iopub.status.idle":"2022-03-16T01:48:34.163033Z","shell.execute_reply.started":"2022-03-16T01:48:26.169661Z","shell.execute_reply":"2022-03-16T01:48:34.16224Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Utils","metadata":{}},{"cell_type":"code","source":"# ====================================================\n# Utils\n# ====================================================\ndef get_score(y_true, y_pred):\n    score = f1_score(y_true, y_pred, average=\"macro\")\n    return score\n\n\ndef init_logger(log_file=OUTPUT_DIR+'train.log'):\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\nLOGGER = init_logger()\n\n\ndef seed_torch(seed=42):\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_torch(seed=CFG.seed)","metadata":{"execution":{"iopub.status.busy":"2022-03-16T01:48:34.164851Z","iopub.execute_input":"2022-03-16T01:48:34.165386Z","iopub.status.idle":"2022-03-16T01:48:34.179283Z","shell.execute_reply.started":"2022-03-16T01:48:34.165343Z","shell.execute_reply":"2022-03-16T01:48:34.178513Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# CV schem","metadata":{}},{"cell_type":"code","source":"%%time\nskf = StratifiedKFold(n_splits=CFG.nfolds, shuffle=True, random_state=CFG.seed)\nfor fold, (trn_idx, vld_idx) in enumerate(skf.split(train, train[CFG.target_col])):\n    train.loc[vld_idx, \"folds\"] = int(fold)\ntrain[\"folds\"] = train[\"folds\"].astype(int)","metadata":{"execution":{"iopub.status.busy":"2022-03-16T01:48:34.181975Z","iopub.execute_input":"2022-03-16T01:48:34.182833Z","iopub.status.idle":"2022-03-16T01:48:42.145067Z","shell.execute_reply.started":"2022-03-16T01:48:34.182793Z","shell.execute_reply":"2022-03-16T01:48:42.144347Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Dataset","metadata":{}},{"cell_type":"code","source":"# ====================================================\n# Dataset\n# ====================================================\nclass TrainDataset(Dataset):\n    def __init__(self, df, transform=None):\n        self.df = df\n        self.file_names = df['directory'].values\n        self.labels = df[CFG.target_col].values\n        self.transform = transform\n        \n    def __len__(self):\n        return len(self.df)\n\n    def __getitem__(self, idx):\n        file_path = self.file_names[idx]\n        try:\n            image = cv2.imread(file_path)\n            image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\n        except:\n            image = Image.open(file_path)\n            image = image.convert(\"RGB\")\n            image = np.array(image)\n        if self.transform:\n            image = self.transform(image=image)['image']\n        label = torch.tensor(self.labels[idx]).float()\n        return image, label\n","metadata":{"execution":{"iopub.status.busy":"2022-03-16T01:48:42.146266Z","iopub.execute_input":"2022-03-16T01:48:42.147051Z","iopub.status.idle":"2022-03-16T01:48:42.156557Z","shell.execute_reply.started":"2022-03-16T01:48:42.147009Z","shell.execute_reply":"2022-03-16T01:48:42.155724Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Transforms","metadata":{}},{"cell_type":"code","source":"# ====================================================\n# Transforms\n# ====================================================\ndef get_transforms(*, data):\n    \n    if data == 'train':\n        return A.Compose(\n        [\n           A.Resize(CFG.size, CFG.size),\n           A.Normalize(\n                mean=[0.485, 0.456, 0.406],\n                std=[0.229, 0.224, 0.225],\n            ),\n            A.Flip(p=0.05),\n            \n            A.Cutout(p=0.05),\n            A.HorizontalFlip(p=0.05),\n            A.VerticalFlip(p=0.05),\n            A.Rotate(limit=180, p=0.05),\n            A.ShiftScaleRotate(\n                shift_limit = 0.1, scale_limit=0.1, rotate_limit=45, p=0.05\n            ),\n            A.HueSaturationValue(\n                hue_shift_limit=0.2, sat_shift_limit=0.4,\n                val_shift_limit=0.2, p=0.05\n            ),\n            A.RandomBrightnessContrast(\n                brightness_limit=(-0.1, 0.1),\n                contrast_limit=(-0.1, 0.1), p=0.05\n            ),\n            ToTensorV2(p=1.0),\n        ]\n    )\n\n    elif data == 'valid':\n        return A.Compose([\n            A.Resize(CFG.size, CFG.size),\n            A.Normalize(\n                mean=[0.485, 0.456, 0.406],\n                std=[0.229, 0.224, 0.225],\n            ),\n            ToTensorV2(),\n        ])","metadata":{"execution":{"iopub.status.busy":"2022-03-16T01:48:42.158149Z","iopub.execute_input":"2022-03-16T01:48:42.15844Z","iopub.status.idle":"2022-03-16T01:48:42.17219Z","shell.execute_reply.started":"2022-03-16T01:48:42.158404Z","shell.execute_reply":"2022-03-16T01:48:42.17135Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train_dataset = TrainDataset(train, transform=get_transforms(data='train'))\n\nfor i in range(5):\n    plt.figure(figsize=(4, 4))\n    image, label = train_dataset[i]\n    plt.imshow(image[0])\n    plt.title(f'label: {label}')\n    plt.show() ","metadata":{"execution":{"iopub.status.busy":"2022-03-16T01:48:42.173518Z","iopub.execute_input":"2022-03-16T01:48:42.173793Z","iopub.status.idle":"2022-03-16T01:48:43.276252Z","shell.execute_reply.started":"2022-03-16T01:48:42.173758Z","shell.execute_reply":"2022-03-16T01:48:43.275597Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Model","metadata":{}},{"cell_type":"code","source":"# ====================================================\n# MODEL\n# ====================================================\nclass CustomModel(nn.Module):\n    def __init__(self, cfg, pretrained=False):\n        super().__init__()\n        self.cfg = cfg\n        self.model = timm.create_model(self.cfg.model_name, pretrained=pretrained, in_chans=3)\n        \n        if cfg.model_name == 'tf_efficientnetv2_b0':\n            self.n_features = self.model.classifier.in_features\n            self.model.classifier = nn.Linear(self.n_features, self.cfg.fc_dim)\n        \n        if cfg.model_name.split('_')[1] == \"efficientnet\":\n            self.n_features = self.model.classifier.in_features\n            self.model.classifier = nn.Linear(self.n_features, self.cfg.fc_dim)\n            \n        if cfg.model_name == 'resnext50_32x4d':\n            self.in_features = self.model.fc.in_features\n            self.model.fc = nn.Linear(self.in_features, self.cfg.fc_dim)\n            \n        if cfg.model_name == 'tresnet_m':\n            self.in_features = self.model.head.fc.in_features\n            self.model.head.fc = nn.Linear(self.in_features, self.cfg.fc_dim)\n            \n        elif cfg.model_name.split('_')[0] == 'vit':\n            self.n_features = self.model.head.in_features\n            self.model.head = nn.Linear(self.n_features, self.cfg.fc_dim)\n        \n        self.fc = nn.Linear(self.cfg.fc_dim, self.cfg.target_size)\n\n    def forward(self, x):\n        features = self.model(x)\n        output = self.fc(features)\n        return output","metadata":{"execution":{"iopub.status.busy":"2022-03-16T01:48:53.828256Z","iopub.execute_input":"2022-03-16T01:48:53.8288Z","iopub.status.idle":"2022-03-16T01:48:53.845101Z","shell.execute_reply.started":"2022-03-16T01:48:53.82876Z","shell.execute_reply":"2022-03-16T01:48:53.844347Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Helper functions","metadata":{}},{"cell_type":"code","source":"# ====================================================\n# Helper functions\n# ====================================================\nclass AverageMeter(object):\n    \"\"\"Computes and stores the average and current value\"\"\"\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\n\n\ndef asMinutes(s):\n    m = math.floor(s / 60)\n    s -= m * 60\n    return '%dm %ds' % (m, s)\n\n\ndef timeSince(since, percent):\n    now = time.time()\n    s = now - since\n    es = s / (percent)\n    rs = es - s\n    return '%s (remain %s)' % (asMinutes(s), asMinutes(rs))\n\n\ndef train_fn(fold, train_loader, model, criterion, optimizer, epoch, scheduler, device):\n    if CFG.apex:\n        scaler = GradScaler()\n    batch_time = AverageMeter()\n    data_time = AverageMeter()\n    losses = AverageMeter()\n    scores = AverageMeter()\n    # switch to train mode\n    model.train()\n    start = end = time.time()\n    global_step = 0\n    for step, (images, labels) in enumerate(train_loader):\n        # measure data loading time\n        data_time.update(time.time() - end)\n        images = images.to(device).float()\n        labels = labels.to(device).long()\n        batch_size = labels.size(0)\n        if CFG.apex:\n            with autocast():\n                y_preds = model(images)\n                loss = criterion(y_preds, labels)\n        else:\n            y_preds = model(images)\n            loss = criterion(y_preds, labels)\n        # record loss\n        losses.update(loss.item(), batch_size)\n        if CFG.gradient_accumulation_steps > 1:\n            loss = loss / CFG.gradient_accumulation_steps\n        if CFG.apex:\n            scaler.scale(loss).backward()\n        else:\n            loss.backward()\n        grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), CFG.max_grad_norm)\n        if (step + 1) % CFG.gradient_accumulation_steps == 0:\n            if CFG.apex:\n                scaler.step(optimizer)\n                scaler.update()\n            else:\n                optimizer.step()\n            optimizer.zero_grad()\n            global_step += 1\n        # measure elapsed time\n        batch_time.update(time.time() - end)\n        end = time.time()\n        if step % CFG.print_freq == 0 or step == (len(train_loader)-1):\n            print('Epoch: [{0}][{1}/{2}] '\n                  'Elapsed {remain:s} '\n                  'Loss: {loss.val:.4f}({loss.avg:.4f}) '\n                  'Grad: {grad_norm:.4f} '\n                  'LR: {lr:.6f}  '\n                  .format(epoch+1, step, len(train_loader), \n                          remain=timeSince(start, float(step+1)/len(train_loader)),\n                          loss=losses,\n                          grad_norm=grad_norm,\n                          lr=scheduler.get_lr()[0]))\n        wandb.log({f\"[fold{fold}] loss\": losses.val,\n                   f\"[fold{fold}] lr\": scheduler.get_lr()[0]})\n    return losses.avg\n\n\ndef valid_fn(valid_loader, model, criterion, device):\n    batch_time = AverageMeter()\n    data_time = AverageMeter()\n    losses = AverageMeter()\n    scores = AverageMeter()\n    # switch to evaluation mode\n    model.eval()\n    preds = []\n    start = end = time.time()\n    for step, (images, labels) in enumerate(valid_loader):\n        # measure data loading time\n        data_time.update(time.time() - end)\n        images = images.to(device).float()\n        labels = labels.to(device).long()\n        batch_size = labels.size(0)\n        # compute loss\n        with torch.no_grad():\n            y_preds = model(images)\n        preds.append(y_preds.argmax(1).to('cpu').numpy())\n        loss = criterion(y_preds, labels)\n        losses.update(loss.item(), batch_size)\n        if CFG.gradient_accumulation_steps > 1:\n            loss = loss / CFG.gradient_accumulation_steps\n        # measure elapsed time\n        batch_time.update(time.time() - end)\n        end = time.time()\n        if step % CFG.print_freq == 0 or step == (len(valid_loader)-1):\n            print('EVAL: [{0}/{1}] '\n                  'Elapsed {remain:s} '\n                  'Loss: {loss.val:.4f}({loss.avg:.4f}) '\n                  .format(step, len(valid_loader),\n                          loss=losses,\n                          remain=timeSince(start, float(step+1)/len(valid_loader))))\n    predictions = np.concatenate(preds)\n    return losses.avg, predictions","metadata":{"execution":{"iopub.status.busy":"2022-03-16T01:48:53.846581Z","iopub.execute_input":"2022-03-16T01:48:53.84709Z","iopub.status.idle":"2022-03-16T01:48:53.876265Z","shell.execute_reply.started":"2022-03-16T01:48:53.847053Z","shell.execute_reply":"2022-03-16T01:48:53.875496Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Train loop","metadata":{}},{"cell_type":"code","source":"# ====================================================\n# Train loop\n# ====================================================\ndef train_loop(folds, fold):\n    \n    LOGGER.info(f\"========== fold: {fold} training ==========\")\n\n    # ====================================================\n    # loader\n    # ====================================================\n    trn_idx = folds[folds['folds'] != fold].index\n    val_idx = folds[folds['folds'] == fold].index\n\n    train_folds = folds.loc[trn_idx].reset_index(drop=True)\n    valid_folds = folds.loc[val_idx].reset_index(drop=True)\n    valid_labels = valid_folds[\"category\"].values\n\n    train_dataset = TrainDataset(train_folds, transform=get_transforms(data='train'))\n    valid_dataset = TrainDataset(valid_folds, transform=get_transforms(data='valid'))\n\n    train_loader = DataLoader(train_dataset,\n                              batch_size=CFG.batch_size, \n                              shuffle=True, \n                              num_workers=CFG.num_workers, pin_memory=True, drop_last=True)\n    valid_loader = DataLoader(valid_dataset, \n                              batch_size=CFG.batch_size * 2, \n                              shuffle=False, \n                              num_workers=CFG.num_workers, pin_memory=True, drop_last=False)\n    \n    # ====================================================\n    # scheduler \n    # ====================================================\n    def get_scheduler(optimizer):\n        if CFG.scheduler=='ReduceLROnPlateau':\n            scheduler = ReduceLROnPlateau(optimizer, **CFG.reduce_params)\n        elif CFG.scheduler=='CosineAnnealingLR':\n            scheduler = CosineAnnealingLR(optimizer, **CFG.cosanneal_params)\n        elif CFG.scheduler=='CosineAnnealingWarmRestarts':\n            scheduler = CosineAnnealingWarmRestarts(optimizer, **CFG.reduce_params)\n        return scheduler\n\n    # ====================================================\n    # model & optimizer\n    # ====================================================\n    model = CustomModel(CFG, pretrained=True)\n    model.to(device)\n\n    optimizer = Adam(model.parameters(), lr=CFG.lr, weight_decay=CFG.weight_decay)\n    scheduler = get_scheduler(optimizer)\n\n    # ====================================================\n    # loop\n    # ====================================================\n    criterion = nn.CrossEntropyLoss()\n    best_score = 0.\n    best_loss = np.inf\n    \n    for epoch in range(CFG.epochs):\n        \n        start_time = time.time()\n        \n        # train\n        avg_loss = train_fn(fold, train_loader, model, criterion, optimizer, epoch, scheduler, device)\n\n        # eval\n        avg_val_loss, preds = valid_fn(valid_loader, model, criterion, device)\n        \n        if isinstance(scheduler, ReduceLROnPlateau):\n            scheduler.step(avg_val_loss)\n        elif isinstance(scheduler, CosineAnnealingLR):\n            scheduler.step()\n        elif isinstance(scheduler, CosineAnnealingWarmRestarts):\n            scheduler.step()\n\n        # scoring\n        \n        #preds_label = np.argmax(preds, axis=1)\n        score = get_score(valid_labels, preds)\n\n        elapsed = time.time() - start_time\n\n        LOGGER.info(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} - Score: {score:.4f}')\n        wandb.log({f\"[fold{fold}] epoch\": epoch+1, \n                   f\"[fold{fold}] avg_train_loss\": avg_loss, \n                   f\"[fold{fold}] avg_val_loss\": avg_val_loss,\n                   f\"[fold{fold}] score\": score})\n\n        if score >= best_score:\n            best_score = score\n            LOGGER.info(f'Epoch {epoch+1} - Save Best Score: {best_score:.4f} Model')\n            torch.save({'model': model.state_dict(), \n                        'preds_score': preds},\n                        OUTPUT_DIR+f'{CFG.model_name}_fold{fold}_best_score.pth')\n            \n        if avg_val_loss < best_loss:\n            best_loss = avg_val_loss\n            LOGGER.info(f'Epoch {epoch+1} - Save Best Loss: {best_loss:.4f} Model')\n            torch.save({'model': model.state_dict(), \n                        'preds_loss': preds},\n                        OUTPUT_DIR+f'{CFG.model_name}_fold{fold}_best_loss.pth')\n        \n        \n    valid_folds[\"preds_score\"] = torch.load(OUTPUT_DIR+f'{CFG.model_name}_fold{fold}_best_score.pth', \n                                      map_location=torch.device('cpu'))['preds_score']\n    valid_folds[\"preds_loss\"] = torch.load(OUTPUT_DIR+f'{CFG.model_name}_fold{fold}_best_loss.pth', \n                                      map_location=torch.device('cpu'))['preds_loss']\n   \n\n    return valid_folds","metadata":{"execution":{"iopub.status.busy":"2022-03-16T01:48:53.877942Z","iopub.execute_input":"2022-03-16T01:48:53.878375Z","iopub.status.idle":"2022-03-16T01:48:53.908683Z","shell.execute_reply.started":"2022-03-16T01:48:53.878331Z","shell.execute_reply":"2022-03-16T01:48:53.907649Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# ====================================================\n# main\n# ====================================================\ndef main():\n\n    \"\"\"\n    Prepare: 1.train \n    \"\"\"\n\n    def get_result(result_df):\n        preds_score = result_df['preds_score'].values\n        preds_loss = result_df['preds_loss'].values\n        labels = result_df[\"category\"].values\n        score = get_score(labels, preds_score)\n        score_loss = get_score(labels, preds_loss)\n        LOGGER.info(f'Score with best score weights: {score:<.4f}')\n        LOGGER.info(f'Score with best loss weights: {score_loss:<.4f}')\n    \n    if CFG.train:\n        # train \n        oof_df = pd.DataFrame()\n        for fold in range(CFG.nfolds):\n            if fold in CFG.trn_folds:\n                _oof_df = train_loop(train, fold)\n                oof_df = pd.concat([oof_df, _oof_df])\n                LOGGER.info(f\"========== fold: {fold} result ==========\")\n                get_result(_oof_df)\n        # CV result\n        LOGGER.info(f\"========== CV ==========\")\n        get_result(oof_df)\n        # save result\n        oof_df.to_csv(OUTPUT_DIR+'oof_df.csv', index=False)\n        \n    wandb.finish()","metadata":{"execution":{"iopub.status.busy":"2022-03-16T01:48:53.910643Z","iopub.execute_input":"2022-03-16T01:48:53.911217Z","iopub.status.idle":"2022-03-16T01:48:53.92297Z","shell.execute_reply.started":"2022-03-16T01:48:53.911179Z","shell.execute_reply":"2022-03-16T01:48:53.922029Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"if __name__ == \"__main__\":\n    main()","metadata":{"execution":{"iopub.status.busy":"2022-03-16T01:48:53.924475Z","iopub.execute_input":"2022-03-16T01:48:53.924974Z"},"trusted":true},"execution_count":null,"outputs":[]}]}