{"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":"
\n

[Pytorch] ArcFace Starter

\n
","metadata":{}},{"cell_type":"markdown","source":"
\n\nIt is based on the work of Vlad Vaduva https://www.kaggle.com/vladvdv/pytorch-train-notebook-arcface-gem-pooling\n \n* v1: initial code for pytorch lightning\n* v2: trying to fix error ...\n* v3: trying to fix error ...\n* v4: trying to fix error ...\n* v5: check 2 epoch training time\n* v6: test half precision(mixed precision) (canceled)\n* v7: test half precision(mixed precision)\n* v5: check 2 epoch mixed precision training time","metadata":{}},{"cell_type":"markdown","source":"#

Install Required Libraries

","metadata":{}},{"cell_type":"code","source":"!pip install timm\n!pip install --upgrade wandb\n!pip install --upgrade pytorch-lightning","metadata":{"_kg_hide-output":true,"execution":{"iopub.status.busy":"2022-02-26T00:25:45.008425Z","iopub.execute_input":"2022-02-26T00:25:45.009163Z","iopub.status.idle":"2022-02-26T00:26:16.303522Z","shell.execute_reply.started":"2022-02-26T00:25:45.009092Z","shell.execute_reply":"2022-02-26T00:26:16.30268Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"#

Import Required Libraries 📚

","metadata":{}},{"cell_type":"code","source":"import os\nimport gc\nimport cv2\nimport math\nimport copy\nimport time\nimport random\n\n# For data manipulation\nimport numpy as np\nimport pandas as pd\n\n# Pytorch Imports\nimport torch\nimport torch.nn as nn\nimport torch.optim as optim\nimport torch.nn.functional as F\nfrom torch.optim import lr_scheduler\nfrom torch.utils.data import Dataset, DataLoader\n\n# Pytorch Lightning Imports\nimport pytorch_lightning as pl\nfrom pytorch_lightning.loggers.wandb import WandbLogger\nfrom pytorch_lightning.callbacks import LearningRateMonitor, ModelCheckpoint\n\n# Utils\nimport joblib\nfrom tqdm import tqdm\nfrom collections import defaultdict\n\n# Sklearn Imports\nfrom sklearn.preprocessing import LabelEncoder\nfrom sklearn.model_selection import StratifiedKFold\n\n# For Image Models\nimport timm\n\n# Albumentations for augmentations\nimport albumentations as A\nfrom albumentations.pytorch import ToTensorV2\n\n# For colored terminal text\nfrom colorama import Fore, Back, Style\nb_ = Fore.BLUE\nsr_ = Style.RESET_ALL\n\nimport warnings\nwarnings.filterwarnings(\"ignore\")\n\n# For descriptive error messages\nos.environ['CUDA_LAUNCH_BLOCKING'] = \"1\"","metadata":{"execution":{"iopub.status.busy":"2022-02-26T00:26:16.305923Z","iopub.execute_input":"2022-02-26T00:26:16.30621Z","iopub.status.idle":"2022-02-26T00:26:26.358769Z","shell.execute_reply.started":"2022-02-26T00:26:16.30617Z","shell.execute_reply":"2022-02-26T00:26:26.357946Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"\"Weights\n\n Weights & Biases (W&B) is a set of machine learning tools that helps you build better models faster. Kaggle competitions require fast-paced model development and evaluation. There are a lot of components: exploring the training data, training different models, combining trained models in different combinations (ensembling), and so on.\n\n> ⏳ Lots of components = Lots of places to go wrong = Lots of time spent debugging\n\nW&B can be useful for Kaggle competition with it's lightweight and interoperable tools:\n\n* Quickly track experiments,
\n* Version and iterate on datasets,
\n* Evaluate model performance,
\n* Reproduce models,
\n* Visualize results and spot regressions,
\n* Share findings with colleagues.\n\nTo learn more about Weights and Biases check out this kernel.","metadata":{}},{"cell_type":"code","source":"import wandb\n\ntry:\n from kaggle_secrets import UserSecretsClient\n user_secrets = UserSecretsClient()\n api_key = user_secrets.get_secret(\"wandb_api\")\n wandb.login(key=api_key)\n anony = None\nexcept:\n anony = \"must\"\n print('If you want to use your W&B account, go to Add-ons -> Secrets and provide your W&B access token. Use the Label name as wandb_api. \\nGet your W&B access token from here: https://wandb.ai/authorize')","metadata":{"execution":{"iopub.status.busy":"2022-02-26T00:26:26.36027Z","iopub.execute_input":"2022-02-26T00:26:26.362518Z","iopub.status.idle":"2022-02-26T00:26:27.666926Z","shell.execute_reply.started":"2022-02-26T00:26:26.362482Z","shell.execute_reply":"2022-02-26T00:26:27.666227Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"#

Training Configuration ⚙️

","metadata":{}},{"cell_type":"code","source":"CONFIG = {\"seed\": 2022,\n \"epochs\": 2,\n \"img_size\": 768,\n \"model_name\": \"tf_efficientnet_b4\",\n \"num_classes\": 15587,\n \"train_batch_size\": 8,\n \"valid_batch_size\": 8,\n \"learning_rate\": 0.0001,\n \"scheduler\": 'OneCycleLR',\n \"min_lr\": 1e-6,\n \"T_max\": 500,\n \"weight_decay\": 1e-6,\n \"n_fold\": 5,\n \"fold_to_run\": 0,\n \"n_accumulate\": 1,\n \"test_mode\":False, # enable for testing pipeline, changes epochs to 2 and uses just 100 training samples\n \"half_precision\": True,\n # ArcFace Hyperparameters\n \"s\": 30.0, \n \"m\": 0.30,\n \"ls_eps\": 0.0,\n \"easy_margin\": False\n }","metadata":{"execution":{"iopub.status.busy":"2022-02-26T00:26:27.670727Z","iopub.execute_input":"2022-02-26T00:26:27.670943Z","iopub.status.idle":"2022-02-26T00:26:27.676477Z","shell.execute_reply.started":"2022-02-26T00:26:27.670917Z","shell.execute_reply":"2022-02-26T00:26:27.675756Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"#

Set Seed for Reproducibility

","metadata":{}},{"cell_type":"code","source":"def set_seed(seed=42):\n '''Sets the seed of the entire notebook so results are the same every time we run.\n This is for REPRODUCIBILITY.'''\n np.random.seed(seed)\n torch.manual_seed(seed)\n torch.cuda.manual_seed(seed)\n # When running on the CuDNN backend, two further options must be set\n torch.backends.cudnn.deterministic = True\n torch.backends.cudnn.benchmark = False\n # Set a fixed value for the hash seed\n os.environ['PYTHONHASHSEED'] = str(seed)\n \nset_seed(CONFIG['seed'])","metadata":{"execution":{"iopub.status.busy":"2022-02-26T00:26:27.677529Z","iopub.execute_input":"2022-02-26T00:26:27.678269Z","iopub.status.idle":"2022-02-26T00:26:27.694396Z","shell.execute_reply.started":"2022-02-26T00:26:27.678232Z","shell.execute_reply":"2022-02-26T00:26:27.693572Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"ROOT_DIR = '../input/happy-whale-and-dolphin'\nTRAIN_DIR = '../input/happy-whale-and-dolphin/train_images'\nTEST_DIR = '../input/happy-whale-and-dolphin/test_images'","metadata":{"execution":{"iopub.status.busy":"2022-02-26T00:26:27.695857Z","iopub.execute_input":"2022-02-26T00:26:27.696441Z","iopub.status.idle":"2022-02-26T00:26:27.700747Z","shell.execute_reply.started":"2022-02-26T00:26:27.696356Z","shell.execute_reply":"2022-02-26T00:26:27.699964Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def get_train_file_path(id):\n return f\"{TRAIN_DIR}/{id}\"","metadata":{"execution":{"iopub.status.busy":"2022-02-26T00:26:27.702455Z","iopub.execute_input":"2022-02-26T00:26:27.703175Z","iopub.status.idle":"2022-02-26T00:26:27.70927Z","shell.execute_reply.started":"2022-02-26T00:26:27.703136Z","shell.execute_reply":"2022-02-26T00:26:27.708495Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"#

Read the Data 📖

","metadata":{}},{"cell_type":"code","source":"df = pd.read_csv(f\"{ROOT_DIR}/train.csv\")\ndf['file_path'] = df['image'].apply(get_train_file_path)\ndf.head()\n\nif CONFIG[\"test_mode\"]==True:\n df=df[:100]\n CONFIG[\"epochs\"] = 2\n CONFIG[\"n_fold\"] = 2\nencoder = LabelEncoder()\ndf['individual_id'] = encoder.fit_transform(df['individual_id'])\n\nwith open(\"le.pkl\", \"wb\") as fp:\n joblib.dump(encoder, fp)","metadata":{"execution":{"iopub.status.busy":"2022-02-26T00:26:27.710589Z","iopub.execute_input":"2022-02-26T00:26:27.711041Z","iopub.status.idle":"2022-02-26T00:26:27.830358Z","shell.execute_reply.started":"2022-02-26T00:26:27.711006Z","shell.execute_reply":"2022-02-26T00:26:27.82934Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"#

Create Folds

","metadata":{}},{"cell_type":"code","source":"skf = StratifiedKFold(n_splits=CONFIG['n_fold'])\n\nfor fold, ( _, val_) in enumerate(skf.split(X=df, y=df.individual_id)):\n df.loc[val_ , \"kfold\"] = fold","metadata":{"execution":{"iopub.status.busy":"2022-02-26T00:26:27.832209Z","iopub.execute_input":"2022-02-26T00:26:27.832602Z","iopub.status.idle":"2022-02-26T00:26:27.843719Z","shell.execute_reply.started":"2022-02-26T00:26:27.832565Z","shell.execute_reply":"2022-02-26T00:26:27.842986Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"#

Dataset Class

","metadata":{}},{"cell_type":"code","source":"class HappyWhaleDataset(Dataset):\n def __init__(self, df, transforms=None):\n self.df = df\n self.file_names = df['file_path'].values\n self.labels = df['individual_id'].values\n self.transforms = transforms\n \n def __len__(self):\n return len(self.df)\n \n def __getitem__(self, index):\n img_path = self.file_names[index]\n img = cv2.imread(img_path)\n img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n label = self.labels[index]\n \n if self.transforms:\n img = self.transforms(image=img)[\"image\"]\n \n return {\n 'image': img,\n 'label': torch.tensor(label, dtype=torch.long)\n }","metadata":{"execution":{"iopub.status.busy":"2022-02-26T00:26:27.847006Z","iopub.execute_input":"2022-02-26T00:26:27.847276Z","iopub.status.idle":"2022-02-26T00:26:27.854646Z","shell.execute_reply.started":"2022-02-26T00:26:27.84724Z","shell.execute_reply":"2022-02-26T00:26:27.85324Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"#

Augmentations

","metadata":{}},{"cell_type":"code","source":"data_transforms = {\n \"train\": A.Compose([\n A.Resize(CONFIG['img_size'], CONFIG['img_size']),\n A.HorizontalFlip(p=0.5),\n A.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 ToTensorV2()], p=1.),\n \n \"valid\": A.Compose([\n A.Resize(CONFIG['img_size'], CONFIG['img_size']),\n A.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 ToTensorV2()], p=1.)\n}","metadata":{"execution":{"iopub.status.busy":"2022-02-26T00:26:27.855791Z","iopub.execute_input":"2022-02-26T00:26:27.85631Z","iopub.status.idle":"2022-02-26T00:26:27.866655Z","shell.execute_reply.started":"2022-02-26T00:26:27.856275Z","shell.execute_reply":"2022-02-26T00:26:27.865955Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"#

GeM Pooling

\n\nCode taken from GeM Pooling Explained\n\n![](https://i.imgur.com/thTgYWG.jpg)","metadata":{}},{"cell_type":"code","source":"class GeM(nn.Module):\n def __init__(self, p=3, eps=1e-6):\n super(GeM, self).__init__()\n self.p = nn.Parameter(torch.ones(1)*p)\n self.eps = eps\n\n def forward(self, x):\n return self.gem(x, p=self.p, eps=self.eps)\n \n def gem(self, x, p=3, eps=1e-6):\n return F.avg_pool2d(x.clamp(min=eps).pow(p), (x.size(-2), x.size(-1))).pow(1./p)\n \n def __repr__(self):\n return self.__class__.__name__ + \\\n '(' + 'p=' + '{:.4f}'.format(self.p.data.tolist()[0]) + \\\n ', ' + 'eps=' + str(self.eps) + ')'","metadata":{"execution":{"iopub.status.busy":"2022-02-26T00:26:27.869722Z","iopub.execute_input":"2022-02-26T00:26:27.869931Z","iopub.status.idle":"2022-02-26T00:26:27.879801Z","shell.execute_reply.started":"2022-02-26T00:26:27.869907Z","shell.execute_reply":"2022-02-26T00:26:27.878753Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"#

ArcFace

\n\nCode taken from Landmark2019-1st-and-3rd-Place-Solution","metadata":{}},{"cell_type":"code","source":"class ArcMarginProduct(nn.Module):\n r\"\"\"Implement of large margin arc distance: :\n Args:\n in_features: size of each input sample\n out_features: size of each output sample\n s: norm of input feature\n m: margin\n cos(theta + m)\n \"\"\"\n def __init__(self, in_features, out_features, s=30.0, \n m=0.50, easy_margin=False, ls_eps=0.0):\n super(ArcMarginProduct, self).__init__()\n self.in_features = in_features\n self.out_features = out_features\n self.s = s\n self.m = m\n self.ls_eps = ls_eps # label smoothing\n self.weight = nn.Parameter(torch.FloatTensor(out_features, in_features))\n nn.init.xavier_uniform_(self.weight)\n\n self.easy_margin = easy_margin\n self.cos_m = math.cos(m)\n self.sin_m = math.sin(m)\n self.th = math.cos(math.pi - m)\n self.mm = math.sin(math.pi - m) * m\n\n def forward(self, input, label, device):\n # --------------------------- cos(theta) & phi(theta) ---------------------\n cosine = F.linear(F.normalize(input), F.normalize(self.weight))\n if(CONFIG['half_precision']==True):\n cosine = cosine.to(torch.float32)\n sine = torch.sqrt(1.0 - torch.pow(cosine, 2))\n phi = cosine * self.cos_m - sine * self.sin_m\n if self.easy_margin:\n phi = torch.where(cosine > 0, phi, cosine)\n else:\n phi = torch.where(cosine > self.th, phi, cosine - self.mm)\n # --------------------------- convert label to one-hot ---------------------\n # one_hot = torch.zeros(cosine.size(), requires_grad=True, device='cuda')\n one_hot = torch.zeros(cosine.size(), device=device)\n one_hot.scatter_(1, label.view(-1, 1).long(), 1)\n if self.ls_eps > 0:\n one_hot = (1 - self.ls_eps) * one_hot + self.ls_eps / self.out_features\n # -------------torch.where(out_i = {x_i if condition_i else y_i) ------------\n output = (one_hot * phi) + ((1.0 - one_hot) * cosine)\n output *= self.s\n\n return output","metadata":{"execution":{"iopub.status.busy":"2022-02-26T00:26:27.882422Z","iopub.execute_input":"2022-02-26T00:26:27.88278Z","iopub.status.idle":"2022-02-26T00:26:27.898078Z","shell.execute_reply.started":"2022-02-26T00:26:27.882739Z","shell.execute_reply":"2022-02-26T00:26:27.897269Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"#

Create Model

","metadata":{}},{"cell_type":"code","source":"class HappyWhaleModel(nn.Module):\n def __init__(self, model_name, pretrained=True):\n super(HappyWhaleModel, self).__init__()\n self.model = timm.create_model(model_name, pretrained=pretrained)\n in_features = self.model.classifier.in_features\n self.model.classifier = nn.Identity()\n self.model.global_pool = nn.Identity()\n self.pooling = GeM()\n self.drop = nn.Dropout(p=0.2, inplace=False)\n self.fc = nn.Linear(in_features,512)\n self.arc = ArcMarginProduct(512, \n CONFIG[\"num_classes\"],\n s=CONFIG[\"s\"], \n m=CONFIG[\"m\"], \n easy_margin=CONFIG[\"ls_eps\"], \n ls_eps=CONFIG[\"ls_eps\"])\n def forward(self, images, labels,device):\n features = self.model(images)\n pooled_features = self.pooling(features).flatten(1)\n pooled_drop = self.drop(pooled_features)\n emb = self.fc(pooled_drop)\n output = self.arc(emb,labels,device)\n return output,emb","metadata":{"execution":{"iopub.status.busy":"2022-02-26T00:26:27.899535Z","iopub.execute_input":"2022-02-26T00:26:27.900097Z","iopub.status.idle":"2022-02-26T00:26:27.912136Z","shell.execute_reply.started":"2022-02-26T00:26:27.900048Z","shell.execute_reply":"2022-02-26T00:26:27.91124Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"#

Create Lightning Model

","metadata":{}},{"cell_type":"code","source":"class LitHappyWhaleModel(pl.LightningModule):\n def __init__(self, model_name, pretrained=True):\n super().__init__()\n self.model = HappyWhaleModel(model_name, pretrained=True)\n self.criterion = nn.CrossEntropyLoss()\n \n # copied from https://pytorchlightning.github.io/lightning-tutorials/notebooks/lightning_examples/text-transformers.html\n def setup(self, stage=None) -> None:\n if stage != \"fit\":\n return\n # Get dataloader by calling it - train_dataloader() is called after setup() by default\n train_loader = self.train_dataloader()\n\n # Calculate total steps\n tb_size = CONFIG['train_batch_size'] * max(1, self.trainer.gpus)\n ab_size = tb_size * self.trainer.accumulate_grad_batches\n self.total_steps = int((len(train_loader.dataset) / ab_size) * float(self.trainer.max_epochs))\n\n \n def forward(self, images, labels):\n return self.model(images, labels, self.device)\n \n \n def training_step(self, batch, batch_idx):\n images = batch['image']\n labels = batch['label']\n \n outputs, emb = self(images, labels)\n loss = self.criterion(outputs, labels)\n \n self.log(\"train/loss\", loss, prog_bar=True)\n return loss\n \n def validation_step(self, batch, batch_idx):\n images = batch['image']\n labels = batch['label']\n \n outputs, emb = self(images, labels)\n loss = self.criterion(outputs, labels)\n \n return loss\n\n def validation_epoch_end(self, validation_step_outputs):\n loss = torch.stack(validation_step_outputs).mean()\n\n self.log(\"val/loss\", loss, prog_bar=True)\n \n\n def configure_optimizers(self):\n optimizer = optim.Adam(self.model.parameters(), lr=CONFIG['learning_rate'], \n weight_decay=CONFIG['weight_decay'])\n\n def fetch_scheduler(optimizer):\n if CONFIG['scheduler'] == 'CosineAnnealingLR':\n scheduler = lr_scheduler.CosineAnnealingLR(optimizer,T_max=CONFIG['T_max'], \n eta_min=CONFIG['min_lr'])\n elif CONFIG['scheduler'] == 'CosineAnnealingWarmRestarts':\n scheduler = lr_scheduler.CosineAnnealingWarmRestarts(optimizer,T_0=CONFIG['T_0'], \n eta_min=CONFIG['min_lr'])\n \n elif CONFIG['scheduler'] == 'OneCycleLR':\n scheduler = lr_scheduler.OneCycleLR(optimizer,max_lr=CONFIG['learning_rate'],total_steps = self.total_steps)\n \n elif CONFIG['scheduler'] == None:\n return None\n \n return scheduler\n \n scheduler = fetch_scheduler(optimizer)\n\n return [optimizer], [scheduler]\n \n def train_dataloader(self):\n df_train = df[df.kfold != CONFIG['fold_to_run']].reset_index(drop=True)\n train_dataset = HappyWhaleDataset(df_train, transforms=data_transforms[\"train\"])\n return DataLoader(train_dataset, batch_size=CONFIG['train_batch_size'], \n num_workers=2, shuffle=True, pin_memory=True, drop_last=True)\n \n def val_dataloader(self):\n df_valid = df[df.kfold == CONFIG['fold_to_run']].reset_index(drop=True)\n valid_dataset = HappyWhaleDataset(df_valid, transforms=data_transforms[\"valid\"])\n return DataLoader(valid_dataset, batch_size=CONFIG['valid_batch_size'], \n num_workers=2, shuffle=False, pin_memory=True)","metadata":{"execution":{"iopub.status.busy":"2022-02-26T00:26:27.915696Z","iopub.execute_input":"2022-02-26T00:26:27.915912Z","iopub.status.idle":"2022-02-26T00:26:27.935327Z","shell.execute_reply.started":"2022-02-26T00:26:27.915875Z","shell.execute_reply":"2022-02-26T00:26:27.934445Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"#

Run Training

","metadata":{}},{"cell_type":"markdown","source":"Start Training","metadata":{}},{"cell_type":"code","source":"checkpoint_callback = ModelCheckpoint(\n monitor=\"val/loss\",\n mode=\"min\",\n filename='{val/loss:.4f}-{epoch}',\n save_top_k=5,\n verbose=True\n)","metadata":{"execution":{"iopub.status.busy":"2022-02-26T00:43:16.980291Z","iopub.execute_input":"2022-02-26T00:43:16.980976Z","iopub.status.idle":"2022-02-26T00:43:16.991288Z","shell.execute_reply.started":"2022-02-26T00:43:16.980938Z","shell.execute_reply":"2022-02-26T00:43:16.990492Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"trainer = pl.Trainer(\n gpus=1,\n logger=WandbLogger(\n project='happywhale', \n config=CONFIG,\n job_type='Train',\n tags=['arcface', 'gem-pooling', 'effnet-b4', '768'],\n anonymous='must'\n ),\n precision=16 if CONFIG['half_precision'] else 32,\n max_epochs = CONFIG['epochs'],\n accumulate_grad_batches = CONFIG['n_accumulate'],\n callbacks=[\n LearningRateMonitor(logging_interval='step'),\n checkpoint_callback\n ]\n)\nmodel = LitHappyWhaleModel(CONFIG['model_name'])\ntrainer.fit(model)","metadata":{"execution":{"iopub.status.busy":"2022-02-26T00:43:18.786461Z","iopub.execute_input":"2022-02-26T00:43:18.786984Z","iopub.status.idle":"2022-02-26T00:43:54.616297Z","shell.execute_reply.started":"2022-02-26T00:43:18.786947Z","shell.execute_reply":"2022-02-26T00:43:54.615474Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"![Upvote!](https://img.shields.io/badge/Upvote-If%20you%20like%20my%20work-07b3c8?style=for-the-badge&logo=kaggle)","metadata":{}}]}