{"cells":[{"metadata":{},"cell_type":"markdown","source":"# Imports, settings and references"},{"metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"cell_type":"code","source":"import os\nimport pandas as pd\nimport numpy as np \n\nfrom PIL import Image\n\nimport torch\nimport torch.nn as nn\nimport torchvision\nfrom torch.utils.data import Dataset\nfrom torchvision import transforms\n\nimport xgboost as xgb\nfrom sklearn.metrics import cohen_kappa_score\n\nimport pickle\n\nDEVICE = torch.device(\"cuda:0\")\nDATA_SOURCE = os.path.join(\"..\",\"input\",\"aptos2019-blindness-detection\")\nDATA_SOURCE = os.path.join(\"..\",\"input\")","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# PyTorch's style data loader defintion\nadapted from : https://www.kaggle.com/abhishek/very-simple-pytorch-training-0-59"},{"metadata":{"_cell_guid":"79c7e3d0-c299-4dcb-8224-4455121ee9b0","_uuid":"d629ff2d2480ee46fbb7e2d37f6b5fab8052498a","trusted":true},"cell_type":"code","source":"class RetinopathyDatasetTrain(Dataset):\n\n    def __init__(self, eval_set=False, random_state=42):\n        # read data list, split in train and eval, select the set\n        csv_file = os.path.join(DATA_SOURCE, \"train.csv\")\n        df = pd.read_csv(csv_file)\n        df_train = df.sample(n=int(df.shape[0]/2), random_state=random_state)\n        ix=[i for i in df.index if i not in df_train.index.values.tolist()]  \n        df_eval = df.loc[ix]            \n        if eval_set : df = df_eval\n        else :        df = df_train\n        self.data = df.reset_index(drop=True)\n            \n    def __len__(self):\n        return len(self.data)\n\n    def __getitem__(self, idx):\n        # get image and process it to tensor ready for the model, extract features\n        folder = os.path.join(DATA_SOURCE, \"train_images\")\n        code = str(self.data.loc[idx, 'id_code'])\n        file = code + \".png\"\n        path = os.path.join(folder, file)\n        imgpil = Image.open(path)\n        base_transforms = transforms.Compose([transforms.Resize((224, 224)),\n                                              transforms.ToTensor(),\n                                              transforms.Normalize([0.485, 0.456, 0.406], \n                                                                   [0.229, 0.224, 0.225])])\n        img_tensor = base_transforms(imgpil)\n        label = self.data.loc[idx, \"diagnosis\"]\n        return {'image': img_tensor, 'labels': label}","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# Extract features"},{"metadata":{"trusted":true},"cell_type":"code","source":"# load the pretrained CNN used as feature extractor\n# no classifier defined, we will take the raw output from the CNN layers\nextractor = torchvision.models.resnet101(pretrained=True)\nextractor.fc = nn.Identity() \nextractor.to(DEVICE)\nextractor.eval()\n\ndata_loader_train = torch.utils.data.DataLoader(RetinopathyDatasetTrain(), \n                            batch_size=64, shuffle=False, num_workers=0, drop_last=False)\ndata_loader_eval = torch.utils.data.DataLoader(RetinopathyDatasetTrain(eval_set=True), \n                            batch_size=64, shuffle=False, num_workers=0, drop_last=False)\n\ndef get_extracted_data(data_loader):\n    for bi, d in enumerate(data_loader):\n        print(\".\", end=\"\")\n        img_tensor = d[\"image\"].to(DEVICE)\n        target = d[\"labels\"].numpy()\n        with torch.no_grad(): feature = extractor(img_tensor)\n        feature = feature.cpu().detach().squeeze(0).numpy()\n        if bi == 0 :\n            features = feature \n            targets = target \n        else :\n            features = np.concatenate([features, feature], axis=0)\n            targets = np.concatenate([targets, target], axis=0)\n    print(\"\")\n    return features, targets\n\nprint(\".............................\")\nfeatures_train, targets_train = get_extracted_data(data_loader_train)\nfeatures_eval, targets_eval = get_extracted_data(data_loader_eval)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# Fit the XGBoost model"},{"metadata":{"trusted":true},"cell_type":"code","source":"XGBOOST_PARAM = {\n    \"random_state\" : 42,\n    'objective': 'multi:softmax',\n    \"num_class\" : 5,\n    \"n_estimators\" : 200,\n    \"eval_metric\" : \"mlogloss\"\n}\n\nxgb_model_1 = xgb.XGBClassifier(**XGBOOST_PARAM)\nxgb_model_1 = xgb_model_1.fit(features_train,targets_train.reshape(-1),\n                        eval_set=[(features_eval, targets_eval.reshape(-1))],\n                        early_stopping_rounds=20,\n                        verbose=False)\nprediction = xgb_model_1.predict(features_eval)\n# pred1 = XGBGBDT.predict_proba(features_eval)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"xgb_model_2 = xgb.XGBClassifier(**XGBOOST_PARAM)\nxgb_model_2 = xgb_model_2.fit(features_eval,targets_eval.reshape(-1),\n                        eval_set=[(features_train, targets_train.reshape(-1))],\n                        early_stopping_rounds=20,\n                        verbose=False)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# Evaluation"},{"metadata":{"trusted":true},"cell_type":"code","source":"print(\"Cohen Kappa quadratic score\", \n      cohen_kappa_score(targets_eval, prediction, weights=\"quadratic\"))\nxgb.plot_importance(xgb_model_1, max_num_features=12)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# Save models for submission"},{"metadata":{"trusted":true},"cell_type":"code","source":"torch.save(extractor.state_dict(), \"resnet101.pth\")\npickle.dump(xgb_model_1, open(\"xgb_model_1\", \"wb\"))\npickle.dump(xgb_model_2, open(\"xgb_model_2\", \"wb\"))","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"","execution_count":null,"outputs":[]}],"metadata":{"kernelspec":{"display_name":"Python 3","language":"python","name":"python3"},"language_info":{"name":"python","version":"3.6.4","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"}},"nbformat":4,"nbformat_minor":1}