{"cells":[{"metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"cell_type":"code","source":"!pip install bbox-visualizer","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"import numpy as np # linear algebra\nimport pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)\nimport os\nfrom sklearn.model_selection import train_test_split\nfrom sklearn import model_selection\n\nimport cv2\nimport matplotlib.pyplot as plt\nimport seaborn as sns\nimport bbox_visualizer as bbv\n\nimport pydicom\nfrom pydicom.pixel_data_handlers.util import apply_voi_lut\nfrom glob import glob\nfrom skimage import exposure\n\nimport torch\n# Neural networks can be constructed using the torch.nn package.\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torch.optim as optim\nfrom torch.utils.data import DataLoader, Dataset\nfrom torch.utils.data.sampler import SequentialSampler\nimport torchvision\n#from torchvision.models.detection import retinanet_resnet50_fpn\n#from torchvision.models.detection.retinanet import RetinaNet\n\nfrom torchvision.models.detection.faster_rcnn import FastRCNNPredictor\nfrom torchvision.models.detection import FasterRCNN\n\nimport albumentations as A\nfrom albumentations.pytorch.transforms import ToTensorV2\n\nimport warnings\n\nwarnings.filterwarnings('ignore')","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"BASE_DIR = '../input/vinbigdata-chest-xray-abnormalities-detection/'\nWEIGHTS_FILE = './'","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"train = pd.read_csv(os.path.join(BASE_DIR, \"train.csv\"))\ntrain.head()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"train = train[train.class_name!='No finding'].reset_index(drop=True)\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"train_df, valid_df = train_test_split(train, test_size=0.33, random_state=42)\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"class LungsAnnotationDataset(Dataset):\n    def __init__(self,dataframe, image_dir, transforms=None):\n        super().__init__()\n        \n        self.image_ids = dataframe['image_id'].unique()\n        self.df = dataframe\n        self.image_dir = image_dir\n        #self.labels = torch.nn.functional.one_hot(torch.tensor(dataframe.class_id))\n        self.transforms = transforms\n    \n    def __getitem__(self, index: int):\n        image_id = self.image_ids[index]\n        records = self.df[self.df['image_id'] == image_id]\n        \n        dcm_data = pydicom.read_file(f'{self.image_dir}/{image_id}.dicom')\n        image = apply_voi_lut(dcm_data.pixel_array, dcm_data)\n        # depending on this value, X-ray may look inverted - fix that:\n        if dcm_data.PhotometricInterpretation == \"MONOCHROME1\":\n            image = np.amax(image) - image\n            \n        #intercept = dcm_data.RescaleIntercept if \"RescaleIntercept\" in dcm_data else 0.0\n        #slope = dcm_data.RescaleSlope if \"RescaleSlope\" in dcm_data else 1.0\n        \n        #if slope != 1:\n        #    image = slope * image.astype(np.float64)\n        #    image = image.astype(np.int16)\n            \n        #image += np.int16(intercept)\n        image = np.stack([image, image, image])\n        image = image - np.min(image)\n\n        image = image / image.max()\n        #image = image * 255.0\n        #image = image.astype('float32')\n        image = exposure.equalize_hist(image) #Normalization of X-ray images\n        image = image.astype('float32')\n\n        image = image.transpose(1,2,0)\n        \n        boxes = records[['x_min','y_min','x_max','y_max']].values\n        #boxes[:, 2] = boxes[:, 0] + boxes[:, 2]\n        #boxes[:, 3] = boxes[:, 1] + boxes[:, 3]\n        \n        area = (boxes[:, 3] - boxes[:, 1]) * (boxes[:, 2] - boxes[:, 0])\n        area = torch.as_tensor(area, dtype=torch.float32)\n        \n        labels = records.class_id.values + 1\n        class_name = records.class_name.values\n        # suppose all instances are not crowd\n        iscrowd = torch.zeros((records.shape[0],), dtype=torch.int64)\n        \n        target = {}\n        target['boxes'] = torch.tensor(boxes)\n        target['labels'] = torch.tensor(labels)\n        #target['class'] = class_name\n        #target['image_id'] = torch.tensor([index])\n        target['area'] = torch.tensor(area)\n        target['iscrowd'] = torch.tensor(iscrowd)\n        \n        if self.transforms:\n            sample = {\n                'image': image,\n                'bboxes': target['boxes'],\n                'labels': labels\n            }\n            sample = self.transforms(**sample)\n            image = sample['image']\n            \n            target['boxes'] = torch.tensor(sample['bboxes'])\n            \n\n        return image, target\n    \n    def __len__(self) -> int:\n        return self.image_ids.shape[0]","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# Albumentations\ndef get_train_transform():\n    return A.Compose([\n        ToTensorV2(),\n    ], bbox_params={'format': 'pascal_voc', 'label_fields': ['labels']})\n\ndef get_valid_transform():\n    return A.Compose([\n        ToTensorV2(),\n    ], bbox_params={'format': 'pascal_voc', 'label_fields': ['labels']})","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"def collate_fn(batch):\n    return tuple(zip(*batch))\n\nDIR_TRAIN = os.path.join(BASE_DIR, \"train\")\ntrain_dataset = LungsAnnotationDataset(train_df, DIR_TRAIN,get_train_transform())\nvalid_dataset = LungsAnnotationDataset(valid_df, DIR_TRAIN,get_valid_transform())","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"train_data_loader = DataLoader(\n    train_dataset,\n    batch_size=6,\n    shuffle=False,\n    num_workers=4,\n    collate_fn=collate_fn\n)\n\nvalid_data_loader = DataLoader(\n    valid_dataset,\n    batch_size=6,\n    shuffle=False,\n    num_workers=4,\n    collate_fn=collate_fn\n)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"images, targets = next(iter(train_data_loader))\nimages = list(image.to(device) for image in images)\n\ntargets = [{k: v.to(device) for k, v in t.items()} for t in targets]","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"print (targets)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"\nnum_classes = 15 \n\nmodel = torchvision.models.detection.fasterrcnn_resnet50_fpn(pretrained=True)\n\nprint(model)\n\nin_features = model.roi_heads.box_predictor.cls_score.in_features\n\nmodel.roi_heads.box_predictor = FastRCNNPredictor(in_features, num_classes)\n\nmodel.load_state_dict(torch.load('../input/train-draft-folding-v0/model.pth', map_location=device))\nmodel.train()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"model.to(device)\nparams = [p for p in model.parameters() if p.requires_grad]\noptimizer = torch.optim.SGD(params, lr=0.0005, momentum=0.9, weight_decay=0.0005)\n#torch.optim.Adam(params, lr = 1e-3)\n#lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=4, gamma=0.1)\nlr_scheduler = None\n\nnum_epochs = 4","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"class Averager:\n    def __init__(self):\n        self.current_total = 0.0\n        self.iterations = 0.0\n\n    def send(self, value):\n        self.current_total += value\n        self.iterations += 1\n\n    @property\n    def value(self):\n        if self.iterations == 0:\n            return 0\n        else:\n            return 1.0 * self.current_total / self.iterations\n\n    def reset(self):\n        self.current_total = 0.0\n        self.iterations = 0.0","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"def train_model(train_dataset, fold):\n    train_data_loader = DataLoader(\n        train_dataset,\n        batch_size=2,\n        shuffle=False,\n        num_workers=4,\n        collate_fn=collate_fn\n    )\n    loss_hist = Averager()\n    itr = 1\n    for epoch in range(num_epochs):\n        loss_hist.reset()\n        model.train()\n        for images, targets in train_data_loader:\n            optimizer.zero_grad()\n            images = list(image.to(device) for image in images)\n            targets = [{k: v.to(device) for k, v in t.items()} for t in targets]\n\n            loss_dict = model(images, targets)\n            #losses = sum(loss for loss in loss_dict.values())\n            loss_classifier, loss_box_reg, loss_objectness, loss_rpn_box_reg = loss_dict.values()\n            losses = sum([loss_objectness, \n                         10*loss_classifier, \n                         10*loss_rpn_box_reg, \n                         (0.5*loss_box_reg**2)\n                         ]\n                        )\n            loss_value = losses.item()\n            loss_hist.send(loss_value)\n\n            losses.backward()\n            optimizer.step()\n            if itr%100==0:\n                print(f\"Fold #{fold} Epoch #{epoch+1} Iteration #{itr} loss: {loss_hist.value}\")\n            \n            itr += 1\n            del loss_dict, loss_classifier, loss_box_reg, loss_objectness, loss_rpn_box_reg,loss_value\n        itr=1    \n        # update the learning rate\n        if lr_scheduler is not None:\n            lr_scheduler.step()\n\n        print(f\"Fold #{fold} Epoch #{epoch+1} loss: {loss_hist.value}\")\n        print(\"Saving Epoch's state...\")\n        torch.save(model.state_dict(), f\"model_state.pth\")","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"k=1\ndf = train.sample(frac=1).reset_index(drop=True)\ny = train.class_id.values\n## GroupK-Fold Splitting\nkfold = model_selection.GroupKFold(n_splits=5)\n\n\nfor train_index, val_index in kfold.split(df, y,groups=df.image_id.values):\n    \n    train_dataset = LungsAnnotationDataset(df.loc[val_index], DIR_TRAIN,get_train_transform())\n    train_model(train_dataset, k)\n    if k==5:\n        valid_dataset = LungsAnnotationDataset(df.loc[train_index], DIR_TRAIN,get_train_transform())\n        val_data_loader = DataLoader(\n                                valid_dataset,\n                                batch_size=6,\n                                shuffle=False,\n                                num_workers=4,\n                                collate_fn=collate_fn\n                            )\n        \n    k+=1","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"torch.save(model.state_dict(), \"model.pth\")\nfrom IPython.display import FileLink\nFileLink(r'model.pth')","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"","execution_count":null,"outputs":[]}],"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":4,"nbformat_minor":4}