{"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":"This is inference notebook for a vanilla approach to the problem: take raw images and run them through a simple network. \nThe training one can be found here: https://www.kaggle.com/konradb/tez-model-train","metadata":{}},{"cell_type":"code","source":"!pip install tez\n!pip install timm","metadata":{"execution":{"iopub.status.busy":"2022-03-23T17:03:48.129352Z","iopub.execute_input":"2022-03-23T17:03:48.129774Z","iopub.status.idle":"2022-03-23T17:04:03.380253Z","shell.execute_reply.started":"2022-03-23T17:03:48.129697Z","shell.execute_reply":"2022-03-23T17:04:03.379385Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import os\nimport albumentations as A\nimport pandas as pd\nimport numpy as np\n\n\nfrom tez import Tez, TezConfig\nfrom tez.callbacks import EarlyStopping\nfrom tez.datasets import ImageDataset\n\nimport torch\nimport torch.nn as nn\nfrom torch.nn import functional as F\n\nfrom sklearn import metrics, model_selection, preprocessing\nimport timm\n\nfrom tqdm import tqdm\n\nfrom sklearn.model_selection import KFold\n\n# ignoring warnings\nimport warnings\nwarnings.simplefilter(\"ignore\")\n\nimport os, cv2, json\nfrom PIL import Image\n\nimport random","metadata":{"_cell_guid":"79c7e3d0-c299-4dcb-8224-4455121ee9b0","_kg_hide-input":true,"_uuid":"d629ff2d2480ee46fbb7e2d37f6b5fab8052498a","papermill":{"duration":5.732608,"end_time":"2021-12-20T22:53:39.01515","exception":false,"start_time":"2021-12-20T22:53:33.282542","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2022-03-23T17:04:03.386542Z","iopub.execute_input":"2022-03-23T17:04:03.38689Z","iopub.status.idle":"2022-03-23T17:04:05.30808Z","shell.execute_reply.started":"2022-03-23T17:04:03.386847Z","shell.execute_reply":"2022-03-23T17:04:05.307222Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class CFG:    \n    # config\n    work_dir = '../input/ultra-mnist/'\n    img_folder = '../input/ultramnist-resized-256/'\n    model_folder = '../input/tez-model-train/'\n    batch_size = 16\n    epochs = 5\n    img_size = 256\n    seed = 42\n    target_size = 28\n    model = 'resnet50'\n    lr = 0.002\n    patience = 4 \n    nfolds = 5","metadata":{"papermill":{"duration":0.026232,"end_time":"2021-12-20T22:53:39.0593","exception":false,"start_time":"2021-12-20T22:53:39.033068","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2022-03-23T17:04:05.309885Z","iopub.execute_input":"2022-03-23T17:04:05.310186Z","iopub.status.idle":"2022-03-23T17:04:05.314514Z","shell.execute_reply.started":"2022-03-23T17:04:05.310143Z","shell.execute_reply":"2022-03-23T17:04:05.31388Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def seed_everything(seed: int = 42) -> None:\n    random.seed(seed)\n    np.random.seed(seed)\n    os.environ[\"PYTHONHASHSEED\"] = str(seed)\n    \nseed_everything(CFG.seed)","metadata":{"execution":{"iopub.status.busy":"2022-03-23T17:04:05.316816Z","iopub.execute_input":"2022-03-23T17:04:05.317434Z","iopub.status.idle":"2022-03-23T17:04:05.328724Z","shell.execute_reply.started":"2022-03-23T17:04:05.317396Z","shell.execute_reply":"2022-03-23T17:04:05.327959Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Functions","metadata":{"papermill":{"duration":0.016944,"end_time":"2021-12-20T22:53:39.093163","exception":false,"start_time":"2021-12-20T22:53:39.076219","status":"completed"},"tags":[]}},{"cell_type":"code","source":"class UModel(nn.Module):\n    def __init__(self, num_classes):\n        super().__init__()\n        self.model = timm.create_model(CFG.model, pretrained= False)\n        n_features = self.model.fc.in_features\n        self.model.fc = nn.Linear(n_features, num_classes)\n\n    def monitor_metrics(self, outputs, targets):\n        device = targets.get_device()\n        outputs = torch.argmax(outputs, dim=1).cpu().detach().numpy()\n        targets = targets.cpu().detach().numpy()\n        f1 = metrics.f1_score(targets, outputs, average=\"macro\")\n        accuracy = metrics.accuracy_score(targets, outputs)\n        return {\"acc\": torch.tensor(accuracy, device=device)}\n\n    def optimizer_scheduler(self):\n        opt = torch.optim.Adam(self.parameters(), lr=1e-3)\n        sch = torch.optim.lr_scheduler.ReduceLROnPlateau(\n            opt,\n            factor=0.5,\n            patience=2,\n            verbose=True,\n            mode=\"max\",\n            threshold=1e-4,\n        )\n        return opt, sch\n\n    def forward(self, image, targets=None):\n        outputs = self.model(image)\n        if targets is not None:\n            loss = nn.CrossEntropyLoss()(outputs, targets)\n            metrics = self.monitor_metrics(outputs, targets)\n            return outputs, loss, metrics\n        return outputs, 0, {}","metadata":{"execution":{"iopub.status.busy":"2022-03-23T17:04:05.330074Z","iopub.execute_input":"2022-03-23T17:04:05.330435Z","iopub.status.idle":"2022-03-23T17:04:05.343469Z","shell.execute_reply.started":"2022-03-23T17:04:05.330391Z","shell.execute_reply":"2022-03-23T17:04:05.342685Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"aug = A.Compose([\n            A.Normalize(\n                mean=[0.5, 0.5, 0.5],\n                std=[0.5, 0.5, 0.5],\n                max_pixel_value=255.0, \n                p=1.0\n            ) ], p=1.)","metadata":{"execution":{"iopub.status.busy":"2022-03-23T17:04:05.344522Z","iopub.execute_input":"2022-03-23T17:04:05.345343Z","iopub.status.idle":"2022-03-23T17:04:05.355141Z","shell.execute_reply.started":"2022-03-23T17:04:05.345291Z","shell.execute_reply":"2022-03-23T17:04:05.354427Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Data\n","metadata":{"papermill":{"duration":0.017345,"end_time":"2021-12-20T22:53:39.269462","exception":false,"start_time":"2021-12-20T22:53:39.252117","status":"completed"},"tags":[]}},{"cell_type":"code","source":"dfx = pd.read_csv(CFG.work_dir + \"train.csv\")\ndfx.rename(columns={\"id\": \"image_id\", \"digit_sum\": \"label\"}, inplace = True)   \ndfx['image_id'] = dfx['image_id'] + '.jpeg'\n\n\n# split into folds\nkf = KFold(n_splits = 5, random_state = 42, shuffle = True)\nfold_id = np.zeros((len(dfx),1))\n\nfor (ii, (train_index, test_index)) in enumerate(kf.split(dfx)):\n    fold_id[test_index] = ii\n    \ndfx['fold'] = fold_id.astype(int)\n\n\ndfx.head()\n","metadata":{"execution":{"iopub.status.busy":"2022-03-23T17:04:05.357523Z","iopub.execute_input":"2022-03-23T17:04:05.35824Z","iopub.status.idle":"2022-03-23T17:04:05.404441Z","shell.execute_reply.started":"2022-03-23T17:04:05.35821Z","shell.execute_reply":"2022-03-23T17:04:05.403699Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# prep test dataset\ndfx_te = pd.read_csv(CFG.work_dir + 'sample_submission.csv')\ntest_image_paths = [CFG.img_folder + 'test_img/' + x + '.jpeg' for x in dfx_te.id.values]\n\n# fake targets\ntest_targets = dfx_te.digit_sum.values\ntest_dataset = ImageDataset(\n    image_paths=test_image_paths,\n    targets= [0] * len(test_image_paths),\n    augmentations = aug,\n)\n","metadata":{"execution":{"iopub.status.busy":"2022-03-23T17:04:05.405538Z","iopub.execute_input":"2022-03-23T17:04:05.405776Z","iopub.status.idle":"2022-03-23T17:04:05.437921Z","shell.execute_reply.started":"2022-03-23T17:04:05.405742Z","shell.execute_reply":"2022-03-23T17:04:05.437279Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# storage for oof and submission\nprval = np.zeros((dfx.shape[0], 28))\nprfull = np.zeros(( len(test_image_paths), 28))","metadata":{"execution":{"iopub.status.busy":"2022-03-23T17:04:05.439Z","iopub.execute_input":"2022-03-23T17:04:05.439319Z","iopub.status.idle":"2022-03-23T17:04:05.443761Z","shell.execute_reply.started":"2022-03-23T17:04:05.439282Z","shell.execute_reply":"2022-03-23T17:04:05.443091Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Model","metadata":{}},{"cell_type":"code","source":"for fold in range(CFG.nfolds):\n    \n    print('----------------------------------------')\n    # split\n    trn_idx = dfx[dfx['fold'] != fold].index\n    val_idx = dfx[dfx['fold'] == fold].index\n    df_train = dfx.loc[trn_idx].reset_index(drop=True)\n    df_valid = dfx.loc[val_idx].reset_index(drop=True)\n    image_path = CFG.img_folder + 'train_img/'\n    train_image_paths = [os.path.join(image_path, x) for x in df_train.image_id.values]\n    valid_image_paths = [os.path.join(image_path, x) for x in df_valid.image_id.values]\n    train_targets = df_train.label.values\n    valid_targets = df_valid.label.values\n \n    valid_dataset = ImageDataset(\n        image_paths=valid_image_paths, targets=valid_targets,\n        augmentations=aug)\n\n    \n    # instantiate and load model\n    n_train_steps = int(len(train_image_paths) / CFG.batch_size * CFG.epochs)\n    model = UModel(num_classes = CFG.target_size) \n    model = Tez(model)\n    config = TezConfig(\n        test_batch_size=64,\n        device=\"cuda\",\n    )\n    \n    model.load(CFG.model_folder + 'model_f' + str(fold) + '.bin',\n               weights_only = True,\n              config = config)\n    \n    print(fold)\n        \n#     # produce predictions - oof \n#     preds_iter = model.predict(valid_dataset, batch_size= 128, n_jobs=-1) \n#     final_preds = []\n#     for preds in preds_iter:\n#         final_preds.append(preds)\n#     final_preds = np.vstack(final_preds)\n#     prval[val_idx,:] = final_preds\n        \n#     print(np.round(np.mean(np.argmax(final_preds, axis=1) == dfx.label[val_idx]),4))\n    \n    # produce predictions - test data\n    preds_iter = model.predict(test_dataset, batch_size= 128, n_jobs=-1) \n    final_preds = []\n    for preds in preds_iter:\n        final_preds.append(preds)\n    final_preds = np.vstack(final_preds)\n\n    prfull += final_preds / CFG.nfolds\n\n    \n# print(np.round(np.mean(np.argmax(prval, axis=1) == dfx.label),4))   \n\n# 2: 0.1998\n# 5: 0.4075","metadata":{"execution":{"iopub.status.busy":"2022-03-23T17:04:05.444905Z","iopub.execute_input":"2022-03-23T17:04:05.445524Z","iopub.status.idle":"2022-03-23T17:06:36.943683Z","shell.execute_reply.started":"2022-03-23T17:04:05.445486Z","shell.execute_reply":"2022-03-23T17:06:36.942831Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Submission","metadata":{}},{"cell_type":"code","source":"final_preds = prfull.argmax(axis = 1)\n\ndfx_te.digit_sum = final_preds\ndfx_te.to_csv(\"submission.csv\", index=False)","metadata":{"trusted":true},"execution_count":null,"outputs":[]}]}