{"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":"

Sartorius: Cell Instance Segmentation

\n \n
","metadata":{}},{"cell_type":"markdown","source":"

Load Modules

","metadata":{}},{"cell_type":"code","source":"# !pip install git+https://github.com/qubvel/segmentation_models.pytorch > /dev/null","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:09.017972Z","iopub.execute_input":"2021-12-24T03:08:09.018413Z","iopub.status.idle":"2021-12-24T03:08:09.026183Z","shell.execute_reply.started":"2021-12-24T03:08:09.018294Z","shell.execute_reply":"2021-12-24T03:08:09.025467Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# import module we'll need to import our custom module\nfrom shutil import copyfile\n\n# copy our file into the working directory (make sure it has .py suffix)\ncopyfile(src = \"../input/losses/Losses.py\", dst = \"../working/Losses.py\")","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:09.028460Z","iopub.execute_input":"2021-12-24T03:08:09.029025Z","iopub.status.idle":"2021-12-24T03:08:09.058507Z","shell.execute_reply.started":"2021-12-24T03:08:09.028990Z","shell.execute_reply":"2021-12-24T03:08:09.057715Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import os\nimport copy\nimport warnings\n\nimport torch\nfrom torch import nn\nfrom torch import optim\nfrom torch.utils.data import Dataset, DataLoader\nfrom torchvision import models\nimport torchvision\nfrom torchvision.transforms import functional as F\n\nfrom torch.utils.tensorboard import SummaryWriter\nimport cv2\n\nimport albumentations as A\nimport albumentations.pytorch\nfrom albumentations import (HorizontalFlip, VerticalFlip, \n ShiftScaleRotate, Normalize, Resize, \n Compose, GaussNoise)\nfrom albumentations.pytorch import ToTensorV2\n\nimport random\nimport numpy as np\nimport pandas as pd\nimport scipy.ndimage as ndi\nfrom tqdm import tqdm, tqdm_notebook\n\nimport pickle\nfrom pathlib import Path\nimport PIL\n\nimport matplotlib.pyplot as plt\nimport matplotlib as mpl\n#import seaborn as sns\nfrom skimage.filters import threshold_otsu\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.metrics import accuracy_score, f1_score\nimport skimage\n\nimport time\nimport collections\nimport itertools\n#import segmentation_models_pytorch as smp\n\nfrom Losses import ComboLoss","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:09.059958Z","iopub.execute_input":"2021-12-24T03:08:09.060576Z","iopub.status.idle":"2021-12-24T03:08:10.779892Z","shell.execute_reply.started":"2021-12-24T03:08:09.060538Z","shell.execute_reply":"2021-12-24T03:08:10.779160Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"os.environ['CUDA_LAUNCH_BLOCKING'] = '1'\nwarnings.filterwarnings(action='ignore', category=DeprecationWarning)\n\n%matplotlib inline\n\ntorch.backends.cudnn.benchmark = True\ntorch.backends.cudnn.enabled = True","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:10.781170Z","iopub.execute_input":"2021-12-24T03:08:10.781415Z","iopub.status.idle":"2021-12-24T03:08:10.787076Z","shell.execute_reply.started":"2021-12-24T03:08:10.781382Z","shell.execute_reply":"2021-12-24T03:08:10.786423Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def set_random(n):\n random.seed(n)\n np.random.seed(n)\n torch.manual_seed(n)\n torch.cuda.manual_seed(n)\n torch.cuda.manual_seed_all(n)\n torch.backends.cudnn.deterministic = True\n os.environ['PYTHONHASHSEED'] = str(n)\n \nset_random(42)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:10.789639Z","iopub.execute_input":"2021-12-24T03:08:10.790423Z","iopub.status.idle":"2021-12-24T03:08:10.802357Z","shell.execute_reply.started":"2021-12-24T03:08:10.790385Z","shell.execute_reply":"2021-12-24T03:08:10.801604Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"!nvidia-smi --query-gpu=gpu_name,driver_version,memory.total --format=csv","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:10.805123Z","iopub.execute_input":"2021-12-24T03:08:10.805330Z","iopub.status.idle":"2021-12-24T03:08:11.486492Z","shell.execute_reply.started":"2021-12-24T03:08:10.805306Z","shell.execute_reply":"2021-12-24T03:08:11.485706Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

Configurations

","metadata":{}},{"cell_type":"code","source":"data_dir = '../input/sartorius-cell-instance-segmentation'\nDEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\nBATCH_SIZE = 4\nNUM_EPOCHS = 30\nTRAIN_CSV = f\"{data_dir}/train.csv\"\nTRAIN_PATH = f\"{data_dir}/train\"\nTEST_PATH = f\"{data_dir}/test\"\nSAMPLE_SUBMISSION = f\"{data_dir}/sample_submission.csv\"\n\nWIDTH = 704\nHEIGHT = 520\n# Threshold for mask length\nTH = 40\nLR = 1e-3\nWEIGHT_DECAY = 0.0005\n\n\n# Normalize to resnet mean and std if True.\nRESNET_MEAN = [0.485, 0.456, 0.406]\nRESNET_STD = [0.229, 0.224, 0.225]\nIMAGE_RESIZE = (224, 224)\n\n\nSAVE_MODEL = True\nLOAD_MODEL = True\n\nWORKERS = 2\n\ncmap = mpl.colors.ListedColormap(['black', 'gray', 'orange', 'green'])","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:11.489078Z","iopub.execute_input":"2021-12-24T03:08:11.489296Z","iopub.status.idle":"2021-12-24T03:08:11.521008Z","shell.execute_reply.started":"2021-12-24T03:08:11.489270Z","shell.execute_reply":"2021-12-24T03:08:11.520311Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

Load the data

","metadata":{}},{"cell_type":"code","source":"train_files = sorted(list(Path(TRAIN_PATH).rglob('*png')))\ntest_files = sorted(list(Path(TEST_PATH).rglob('*.png')))\nprint(f'Number of pictures in train dir: {len(train_files)} pcs')\nprint()\nprint(f'Number of pictures in test dir: {len(test_files)} pcs')","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:11.522189Z","iopub.execute_input":"2021-12-24T03:08:11.522576Z","iopub.status.idle":"2021-12-24T03:08:11.551946Z","shell.execute_reply.started":"2021-12-24T03:08:11.522538Z","shell.execute_reply":"2021-12-24T03:08:11.551153Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"cell_df = pd.read_csv(TRAIN_CSV)\ncell_df.info()\ncell_df.sample(4)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:11.553328Z","iopub.execute_input":"2021-12-24T03:08:11.553575Z","iopub.status.idle":"2021-12-24T03:08:11.906698Z","shell.execute_reply.started":"2021-12-24T03:08:11.553542Z","shell.execute_reply":"2021-12-24T03:08:11.905943Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

Utilits

","metadata":{}},{"cell_type":"markdown","source":"***Run Length Encoding: Annotation Data to Mask Image¶***\n\nNext funtion is refferred to INVERSION's notebook'Run Length Decoding - Quick Start'.\n\nhttps://www.kaggle.com/inversion/run-length-decoding-quick-start","metadata":{}},{"cell_type":"markdown","source":"***What is Run-length encoding (RLE)?***\n\nRun-length encoding (RLE) is a form of lossless data compression in which runs of data (sequences in which the same data value occurs in many consecutive data elements) are stored as a single data value and count, rather than as the original run. This is most efficient on data that contains many such runs, for example, simple graphic images such as icons, line drawings, Conway's Game of Life, and animations. For files that do not have many runs, RLE could increase the file size. (from Wikipedia)\n\n#### [WIKI](https://en.wikipedia.org/wiki/Run-length_encoding)","metadata":{}},{"cell_type":"code","source":"def rle_decode(mask_rle,shape,color=1):\n '''\n mask_rle: run-length as string formated (start length)\n shape: (height,width) of array to return \n if colors = 1, then:\n Returns numpy array, 1 - mask, 0 - background\n '''\n \n s = mask_rle.split()\n # get starting pixel and cells's length\n start, length = [np.asarray(x, dtype=int) for x in (s[0:][::2], s[1:][::2])]\n start -=1\n # Ending pixels\n ends = start + length\n img = np.zeros((shape[0] * shape[1], shape [2]), dtype=np.float32)\n for lo, hi in zip (start, ends):\n img[lo : hi] = color\n return img.reshape(shape)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:11.908039Z","iopub.execute_input":"2021-12-24T03:08:11.908306Z","iopub.status.idle":"2021-12-24T03:08:11.918151Z","shell.execute_reply.started":"2021-12-24T03:08:11.908275Z","shell.execute_reply":"2021-12-24T03:08:11.917402Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Masks have an abnormally long continuous lines.\n\nThe mask is not representative enough.\n\nWe can take only correct masks to train our model.\n\n**Thanks** [Slavko Prytula](https://www.kaggle.com/slavkoprytula/mask-correction-mask-filtering-updated#Utils)","metadata":{}},{"cell_type":"code","source":"def clean_mask(mask):\n '''\n Function is called to identify whether the mask is broken\n returns True or False state and also returns a mask\n '''\n mask = mask > threshold_otsu(np.array(mask).astype(np.uint8))\n mask = ndi.binary_fill_holes(mask).astype(np.uint8)\n \n # New code for mask acceptance\n contours, hierarchy = cv2.findContours(mask, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)\n c = contours[0][:, 0]\n diff = c - np.roll(c, 1, 0)\n # find horizontal lines longer than threshold\n targets = (diff[:, 1] == 0) & (np.abs(diff[:, 0]) >= TH) \n \n return mask, (True in targets)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:11.919551Z","iopub.execute_input":"2021-12-24T03:08:11.919831Z","iopub.status.idle":"2021-12-24T03:08:11.928365Z","shell.execute_reply.started":"2021-12-24T03:08:11.919797Z","shell.execute_reply":"2021-12-24T03:08:11.927689Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def build_mask(labels, input_shape, colors=True):\n height, width = input_shape\n masks = np.zeros((height, width, 1))\n #masks = np.zeros((height,width), dtype=np.uint8)\n for i, label in enumerate(labels):\n a_mask = rle_decode(label, shape=(height,width, 1))\n a_mask = np.array(a_mask) > 0 \n a_mask, broken_mask = clean_mask(a_mask)\n if broken_mask:\n continue\n masks += a_mask\n masks = masks.clip(0,1)\n return masks","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:11.930450Z","iopub.execute_input":"2021-12-24T03:08:11.931398Z","iopub.status.idle":"2021-12-24T03:08:11.940890Z","shell.execute_reply.started":"2021-12-24T03:08:11.931361Z","shell.execute_reply":"2021-12-24T03:08:11.940074Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

Visualize the Data

","metadata":{}},{"cell_type":"code","source":"temp = cell_df.groupby(['id','cell_type'])['annotation'].agg(lambda x: list(x)).reset_index()\nfor i in range(4):\n #rand_idx = int(np.random.uniform(0, 607))\n #im = PIL.Image.open(train_files[rand_idx])\n im = PIL.Image.open(train_files[i])\n \n start = time.time()\n label = (train_files[i]).stem\n \n cell_type = temp[temp['id'] == label]\n cell_type = str(cell_type['cell_type'].values)\n print(\"--- %s seconds in label ---\" % (time.time() - start))\n \n start = time.time()\n sample_im_df = cell_df[cell_df['id'] == label]\n print(\"--- %s seconds in sorting df ---\" % (time.time() - start))\n \n sample_rles = sample_im_df['annotation'].values\n start = time.time()\n sample_masks1 = build_mask(sample_rles,input_shape=(HEIGHT,WIDTH), colors=False)\n print(\"--- %s seconds in mask1 ---\" % (time.time() - start))\n \n start = time.time()\n sample_masks2 = build_mask(sample_rles,input_shape=(HEIGHT,WIDTH), colors=True)\n print(\"--- %s seconds in mask2 ---\" % (time.time() - start))\n \n \n\n fig, (ax1, ax2, ax3) = plt.subplots(1, 3, figsize=(30, 20),)\n ax1.imshow(im,cmap='gray')\n ax2.imshow(sample_masks1, cmap='gray')\n ax3.imshow(im,cmap='gray')\n ax3.imshow(sample_masks2,alpha = .2)\n \n ax1.set_title('Cell id: ' + str(label) + 'Type:' + str(cell_type))\n ax2.set_title('Mask for id: ' + str(label) + 'Type:' + str(cell_type))\n ax3.set_title('Mask & pic: ' + str(label) + 'Type:' + str(cell_type))\n plt.show();\ndel temp, im, sample_masks1, sample_masks2","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:11.943252Z","iopub.execute_input":"2021-12-24T03:08:11.943760Z","iopub.status.idle":"2021-12-24T03:08:42.283921Z","shell.execute_reply.started":"2021-12-24T03:08:11.943715Z","shell.execute_reply":"2021-12-24T03:08:42.283252Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

Preparing DataSets And DataLoaders

","metadata":{}},{"cell_type":"markdown","source":"

Define Augmenting

","metadata":{}},{"cell_type":"code","source":"def get_augmentation(pic_size,mode):\n if mode !='train':\n augmentation_test = albumentations.Compose([A.Resize(pic_size,pic_size),\n A.Normalize(\n mean=RESNET_MEAN,\n std=RESNET_STD),\n A.pytorch.transforms.ToTensorV2()])\n return augmentation_test\n else:\n augmentation_train = A.Compose([\n A.Resize(pic_size,pic_size),\n A.RandomCrop(pic_size, pic_size),\n A.OneOf([\n A.HorizontalFlip(p=0.5),\n A.RandomRotate90(p=0.5),\n #A.VerticalFlip(p=0.5)\n ], p=0.5),\n A.OneOf([\n A.GaussNoise(),], p=0.8),\n\n A.OneOf([\n A.CLAHE(clip_limit=2),\n A.RandomBrightnessContrast(),], p=0.2),\n# A.HueSaturationValue(p=0.2),\n A.Normalize(\n mean=RESNET_MEAN,\n std=RESNET_STD),\n A.pytorch.transforms.ToTensorV2()], p=1)\n return augmentation_train","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:42.287385Z","iopub.execute_input":"2021-12-24T03:08:42.288807Z","iopub.status.idle":"2021-12-24T03:08:42.298633Z","shell.execute_reply.started":"2021-12-24T03:08:42.288768Z","shell.execute_reply":"2021-12-24T03:08:42.297808Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class NeuroDataSet(Dataset):\n def __init__(self, files, df: pd.core.frame.DataFrame,train:bool, transform=None): \n self.files = files\n self.df = df\n self.height = HEIGHT\n self.width = WIDTH\n self.gb = self.df.groupby('id')\n self.transform = transform\n self.len_ = len(self.files)\n self.train = train\n \n \n def load_sample(self, file):\n img = cv2.imread(str(file), cv2.IMREAD_COLOR)\n img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n img_id = file.stem\n return img, img_id\n \n def __getitem__(self,index):\n # load image\n img, img_id = self.load_sample(self.files[index])\n df_temp = self.gb.get_group(img_id)\n \n annotations = df_temp['annotation'].values\n mask = build_mask(annotations, (self.height,self.width), colors=False)\n mask = mask.astype('float32')\n if self.transform is not None:\n augmented = self.transform(image = img, mask = mask)\n img = augmented['image']\n mask = augmented['mask']\n return img, mask.reshape((1, IMAGE_RESIZE[0], IMAGE_RESIZE[1]))\n else:\n return img, mask.reshape((1, IMAGE_RESIZE[0], IMAGE_RESIZE[1]))\n \n def __len__(self):\n return self.len_","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:42.300461Z","iopub.execute_input":"2021-12-24T03:08:42.300964Z","iopub.status.idle":"2021-12-24T03:08:42.317079Z","shell.execute_reply.started":"2021-12-24T03:08:42.300926Z","shell.execute_reply":"2021-12-24T03:08:42.316357Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

Split train-val set

","metadata":{}},{"cell_type":"code","source":"# Split dataset to train and val sets\ntrain_pics, val_pics = train_test_split(train_files, test_size=0.1)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:42.318516Z","iopub.execute_input":"2021-12-24T03:08:42.318981Z","iopub.status.idle":"2021-12-24T03:08:42.329428Z","shell.execute_reply.started":"2021-12-24T03:08:42.318944Z","shell.execute_reply":"2021-12-24T03:08:42.328712Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Extract pic Id from Path\ntrain_pic_id = [path.stem for path in train_pics]\nval_pic_id = [path.stem for path in val_pics]\n# Print number of files in tran and val sets\nprint(f'Number of pictures in train set: {len(train_pics)}')\nprint()\nprint(f'Number of pictures in val set: {len(val_pics)}')","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:42.330870Z","iopub.execute_input":"2021-12-24T03:08:42.331342Z","iopub.status.idle":"2021-12-24T03:08:42.362891Z","shell.execute_reply.started":"2021-12-24T03:08:42.331307Z","shell.execute_reply":"2021-12-24T03:08:42.359640Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# split pandas dataframe with annotations for train df and val df\n# in consistance with given file Ids\ndf_train = cell_df[(cell_df['id'].isin(train_pic_id))]\ndf_val = cell_df[(cell_df['id'].isin(val_pic_id))]\n\nprint(f\"Number of picture's id in train set {df_train.id.nunique()}\")\nprint()\nprint(f\"Number of picture's id in val set {df_val.id.nunique()}\")","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:42.363911Z","iopub.execute_input":"2021-12-24T03:08:42.364167Z","iopub.status.idle":"2021-12-24T03:08:42.424825Z","shell.execute_reply.started":"2021-12-24T03:08:42.364132Z","shell.execute_reply":"2021-12-24T03:08:42.420640Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train_dataset = NeuroDataSet(train_pics,\n df_train,\n train = True,\n transform=get_augmentation(224, 'train')\n )","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:42.427029Z","iopub.execute_input":"2021-12-24T03:08:42.427454Z","iopub.status.idle":"2021-12-24T03:08:42.437503Z","shell.execute_reply.started":"2021-12-24T03:08:42.427384Z","shell.execute_reply":"2021-12-24T03:08:42.434684Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train_loader = DataLoader(train_dataset,\n batch_size=BATCH_SIZE,\n shuffle=False,\n pin_memory=True,\n num_workers=WORKERS,)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:42.439351Z","iopub.execute_input":"2021-12-24T03:08:42.439912Z","iopub.status.idle":"2021-12-24T03:08:42.455889Z","shell.execute_reply.started":"2021-12-24T03:08:42.439877Z","shell.execute_reply":"2021-12-24T03:08:42.454547Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"val_dataset = NeuroDataSet(val_pics,\n df_val,train=False,\n transform=get_augmentation(224, 'val')\n )","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:42.457054Z","iopub.execute_input":"2021-12-24T03:08:42.457602Z","iopub.status.idle":"2021-12-24T03:08:42.464048Z","shell.execute_reply.started":"2021-12-24T03:08:42.457563Z","shell.execute_reply":"2021-12-24T03:08:42.463074Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"val_loader = DataLoader(val_dataset,\n batch_size=BATCH_SIZE,\n shuffle=False,\n pin_memory=True,\n num_workers=WORKERS,)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:42.465470Z","iopub.execute_input":"2021-12-24T03:08:42.466278Z","iopub.status.idle":"2021-12-24T03:08:42.474026Z","shell.execute_reply.started":"2021-12-24T03:08:42.466207Z","shell.execute_reply":"2021-12-24T03:08:42.473082Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"batch = next(iter(train_loader)) \nimages, masks = batch\nprint(f\"image shape: {images.shape},\\nmask shape:{masks.shape},\\nbatch len: {len(batch)}\")\ndel batch, images, masks","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:08:42.476902Z","iopub.execute_input":"2021-12-24T03:08:42.477788Z","iopub.status.idle":"2021-12-24T03:09:06.058822Z","shell.execute_reply.started":"2021-12-24T03:08:42.477742Z","shell.execute_reply":"2021-12-24T03:09:06.057978Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

Visualize Augmented DataSet

","metadata":{}},{"cell_type":"code","source":"def visualize_augmentations(dataset, idx=1, samples=9, cols=3):\n dataset = copy.deepcopy(dataset)\n dataset.transform = A.Compose([t for t in dataset.transform if not isinstance(t, (A.Normalize, A.pytorch.ToTensorV2))])\n rows = samples // cols\n figure, ax = plt.subplots(nrows=rows, ncols=cols, figsize=(12, 12))\n for i in range(samples):\n image, mask = dataset[idx]\n ax.ravel()[i].imshow(image)\n ax.ravel()[i].set_axis_off()\n plt.tight_layout()\n plt.show()\n del dataset","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:06.060182Z","iopub.execute_input":"2021-12-24T03:09:06.060807Z","iopub.status.idle":"2021-12-24T03:09:06.067697Z","shell.execute_reply.started":"2021-12-24T03:09:06.060764Z","shell.execute_reply":"2021-12-24T03:09:06.066966Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"visualize_augmentations(train_dataset,idx=3)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:06.068916Z","iopub.execute_input":"2021-12-24T03:09:06.069326Z","iopub.status.idle":"2021-12-24T03:09:10.351089Z","shell.execute_reply.started":"2021-12-24T03:09:06.069289Z","shell.execute_reply":"2021-12-24T03:09:10.350478Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

UNET from SCRATCH

","metadata":{}},{"cell_type":"code","source":"# Create convolution block class\nclass Conv_Block(nn.Module):\n '''convolution ==> BN ==> ReLU'''\n \n def __init__(self, in_channels, out_channels):\n super().__init__()\n self.conv = nn.Sequential(\n nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1),\n nn.BatchNorm2d(out_channels),\n nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),\n nn.BatchNorm2d(out_channels),\n nn.ReLU(inplace=True)\n )\n \n def forward(self, x):\n x = self.conv(x)\n return x","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:10.352396Z","iopub.execute_input":"2021-12-24T03:09:10.352812Z","iopub.status.idle":"2021-12-24T03:09:10.360016Z","shell.execute_reply.started":"2021-12-24T03:09:10.352778Z","shell.execute_reply":"2021-12-24T03:09:10.359434Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class Encoder(nn.Module):\n \"\"\"Downscaling with maxpool then double conv\"\"\"\n\n def __init__(self, in_channels, out_channels):\n super().__init__()\n self.maxpool_conv = nn.Sequential(\n Conv_Block(in_channels, out_channels),\n nn.MaxPool2d(2)\n )\n\n def forward(self, x):\n x = x.cuda()\n return self.maxpool_conv(x)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:10.361081Z","iopub.execute_input":"2021-12-24T03:09:10.361764Z","iopub.status.idle":"2021-12-24T03:09:10.372892Z","shell.execute_reply.started":"2021-12-24T03:09:10.361722Z","shell.execute_reply":"2021-12-24T03:09:10.372251Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class Decoder(nn.Module):\n \"\"\"Upscaling then double conv\"\"\"\n def __init__(self, in_channels, out_channels, bilinear=True):\n super().__init__()\n # if bilinear, use the normal convolutions to reduce the number of channels\n if bilinear:\n self.up = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)\n else:\n self.up = nn.ConvTranspose2d(in_channels // 2, in_channels // 2, kernel_size=2, stride=2)\n \n self.conv = Conv_Block(in_channels, out_channels)\n \n def forward(self, x1, x2):\n x1 = x1.cuda()\n x2 = x2.cuda()\n x1 = self.up(x1)\n x = torch.cat([x2, x1], dim=1)\n x = self.conv(x)\n return x","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:10.374178Z","iopub.execute_input":"2021-12-24T03:09:10.374612Z","iopub.status.idle":"2021-12-24T03:09:10.384553Z","shell.execute_reply.started":"2021-12-24T03:09:10.374579Z","shell.execute_reply":"2021-12-24T03:09:10.383802Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class OutConv(nn.Module):\n def __init__(self, in_channels, out_channels):\n super(OutConv, self).__init__()\n self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=1)\n\n def forward(self, x):\n return self.conv(x)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:10.385849Z","iopub.execute_input":"2021-12-24T03:09:10.386292Z","iopub.status.idle":"2021-12-24T03:09:10.392829Z","shell.execute_reply.started":"2021-12-24T03:09:10.386259Z","shell.execute_reply":"2021-12-24T03:09:10.392060Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class UNet(nn.Module):\n def __init__(self, n_channels, n_classes, bilinear=True):\n super(UNet, self).__init__()\n self.n_channels = n_channels\n self.n_classes = n_classes\n self.bilinear = bilinear\n \n self.inc = Conv_Block(n_channels, 64)\n self.enc1 = Encoder(64, 128)\n self.enc2 = Encoder(128, 256)\n self.enc3 = Encoder(256, 512)\n self.enc4 = Encoder(512, 512)\n self.dec1 = Decoder(1024, 256, bilinear)\n self.dec2 = Decoder(512, 128, bilinear)\n self.dec3 = Decoder(256, 64, bilinear)\n self.dec4 = Decoder(128, 64, bilinear)\n self.outc = OutConv(64, n_classes)\n \n def forward(self, x):\n x1 = self.inc(x)\n x2 = self.enc1(x1)\n x3 = self.enc2(x2)\n x4 = self.enc3(x3)\n x5 = self.enc4(x4)\n x = self.dec1(x5, x4)\n x = self.dec2(x, x3)\n x = self.dec3(x, x2)\n x = self.dec4(x, x1)\n logits = self.outc(x)\n return logits","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:10.394107Z","iopub.execute_input":"2021-12-24T03:09:10.394565Z","iopub.status.idle":"2021-12-24T03:09:10.407739Z","shell.execute_reply.started":"2021-12-24T03:09:10.394533Z","shell.execute_reply":"2021-12-24T03:09:10.407092Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

Define loss functions

","metadata":{}},{"cell_type":"code","source":"from Losses import ComboLoss\n# JaccardLoss = smp.losses.JaccardLoss(mode='binary') \n# Dice = smp.losses.DiceLoss(mode='binary',)\n# BCELoss = smp.losses.SoftBCEWithLogitsLoss()","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:10.409121Z","iopub.execute_input":"2021-12-24T03:09:10.409588Z","iopub.status.idle":"2021-12-24T03:09:10.422343Z","shell.execute_reply.started":"2021-12-24T03:09:10.409551Z","shell.execute_reply":"2021-12-24T03:09:10.421410Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

Define Save Weights function

","metadata":{}},{"cell_type":"code","source":"def save_checkpoint(model, optimizer, filename=\"my_checkpoint.pth.tar\"):\n print(\"=> Saving checkpoint\")\n checkpoint = {\n \"state_dict\": model.state_dict(),\n \"optimizer\": optimizer.state_dict(),\n }\n torch.save(checkpoint, filename)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:10.424002Z","iopub.execute_input":"2021-12-24T03:09:10.424501Z","iopub.status.idle":"2021-12-24T03:09:10.430997Z","shell.execute_reply.started":"2021-12-24T03:09:10.424464Z","shell.execute_reply":"2021-12-24T03:09:10.430220Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def load_checkpoint(checkpoint_file, model, optimizer, lr):\n print(\"=> Loading checkpoint\")\n checkpoint = torch.load(checkpoint_file, map_location=\"cuda\")\n model.load_state_dict(checkpoint[\"state_dict\"])\n optimizer.load_state_dict(checkpoint[\"optimizer\"])\n\n # If we don't do this then it will just have learning rate of old checkpoint\n # and it will lead to many hours of debugging \\:\n for param_group in optimizer.param_groups:\n param_group[\"lr\"] = lr","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:10.432441Z","iopub.execute_input":"2021-12-24T03:09:10.433073Z","iopub.status.idle":"2021-12-24T03:09:10.439778Z","shell.execute_reply.started":"2021-12-24T03:09:10.433001Z","shell.execute_reply":"2021-12-24T03:09:10.439083Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

Train

","metadata":{}},{"cell_type":"code","source":"def train_loop(model, optimizer, criterion, train_loader, device=DEVICE):\n running_loss = 0\n model.train()\n pbar = tqdm(train_loader, desc='Iterating over train data')\n for imgs, masks in pbar:\n # pass to device\n imgs = imgs.to(device)\n masks = masks.to(device)\n # forward\n out = model(imgs)\n loss = criterion(out, masks)\n running_loss += loss.item()*imgs.shape[0] # += loss * current batch size\n # optimize\n optimizer.zero_grad()\n loss.backward()\n optimizer.step()\n running_loss /= len(train_loader.sampler)\n del imgs, masks\n torch.cuda.empty_cache()\n return running_loss","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:10.441270Z","iopub.execute_input":"2021-12-24T03:09:10.441860Z","iopub.status.idle":"2021-12-24T03:09:10.450392Z","shell.execute_reply.started":"2021-12-24T03:09:10.441821Z","shell.execute_reply":"2021-12-24T03:09:10.449581Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"@torch.no_grad()\ndef eval_loop(model, criterion, eval_loader, device=DEVICE):\n running_loss = 0\n model.eval()\n with torch.no_grad():\n accuracy, f1_scores = [], []\n pbar = tqdm(eval_loader, desc='Iterating over evaluation data')\n for imgs, masks in pbar:\n # pass to device\n imgs = imgs.to(device)\n masks = masks.to(device)\n # forward\n out = model(imgs)\n loss = criterion(out, masks)\n running_loss += loss.item()*imgs.shape[0]\n # calculate predictions using output\n predicted = (out > 0.5).float()\n predicted = predicted.view(-1).cpu().numpy()\n labels = masks.view(-1).cpu().numpy()\n accuracy.append(accuracy_score(labels, predicted))\n f1_scores.append(f1_score(labels, predicted))\n acc = sum(accuracy)/len(accuracy)\n f1 = sum(f1_scores)/len(f1_scores)\n running_loss /= len(eval_loader.sampler)\n return {\n 'accuracy':acc,\n 'f1_macro':f1, \n 'loss':running_loss}","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:10.451747Z","iopub.execute_input":"2021-12-24T03:09:10.452053Z","iopub.status.idle":"2021-12-24T03:09:10.463108Z","shell.execute_reply.started":"2021-12-24T03:09:10.452018Z","shell.execute_reply":"2021-12-24T03:09:10.462406Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"@torch.no_grad()\ndef visualize_train(model,val_loader, device=DEVICE):\n model.eval()\n X_val, Y_val = next(iter(val_loader))\n Y_hat = model(X_val.to(device))\n \n Y_hat = torch.tanh(Y_hat)\n Y_hat = Y_hat.detach().cpu().numpy()\n X_val = X_val.detach().cpu().numpy()\n \n #clear_output(wait=True)\n for k in range(2):\n fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(10, 5),)\n ax1.imshow(np.rollaxis(Y_val[k].numpy(), 0, 3),cmap='gray')\n ax2.imshow(((Y_hat[k, 0] * 255)), cmap='gray')\n ax1.set_axis_off()\n ax2.set_axis_off()\n ax1.set_title('Real')\n ax2.set_title('Output')\n plt.show()\n del X_val, Y_hat","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:10.465564Z","iopub.execute_input":"2021-12-24T03:09:10.466506Z","iopub.status.idle":"2021-12-24T03:09:10.474811Z","shell.execute_reply.started":"2021-12-24T03:09:10.466469Z","shell.execute_reply":"2021-12-24T03:09:10.473966Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def train(model, optimizer, criterion, train_loader, valid_loader,\n device=DEVICE, \n num_epochs=NUM_EPOCHS, \n valid_loss_min=np.inf,\n vis=False,\n model_name=None):\n #step = 0\n history = []\n \n for e in range(num_epochs):\n # train for epoch\n train_loss = train_loop(\n model, optimizer, criterion, train_loader, device=device)\n torch.cuda.empty_cache() #!!!!!!!!!!!!!!!!!\n # evaluate on validation set\n metrics = eval_loop(\n model, criterion, valid_loader, device=device\n )\n history.append((train_loss,metrics[\"loss\"],metrics[\"accuracy\"],metrics[\"f1_macro\"]))\n \n #writer = SummaryWriter(f\"runs/Sartorius LR_{LR}_BS_{BATCH_SIZE}\")\n # visualize model in tensorboard\n images, masks = next(iter(train_loader))\n #writer.add_graph(model, images.to(device))\n #writer.close()\n \n if vis:\n visualize_train(model,val_loader,device=DEVICE)\n \n # show progress\n print_string = f'Epoch: {e+1} '\n print_string+= f'TrainLoss: {train_loss:.5f} '\n print_string+= f'ValidLoss: {metrics[\"loss\"]:.5f} '\n print_string+= f'ACC: {metrics[\"accuracy\"]:.5f} '\n print_string+= f'F1: {metrics[\"f1_macro\"]:.3f}'\n print(print_string)\n\n\n # save the model \n if SAVE_MODEL and metrics[\"loss\"] <= valid_loss_min:\n save_checkpoint(model, optimizer, filename=model_name)\n valid_loss_min = metrics[\"loss\"]\n\n return history","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:10.476032Z","iopub.execute_input":"2021-12-24T03:09:10.476510Z","iopub.status.idle":"2021-12-24T03:09:10.487502Z","shell.execute_reply.started":"2021-12-24T03:09:10.476473Z","shell.execute_reply":"2021-12-24T03:09:10.486688Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"model = UNet(3, 1).to(DEVICE)\noptimizer = torch.optim.Adam(model.parameters(), lr=LR, weight_decay=WEIGHT_DECAY )\ncriterion = ComboLoss(**{'weights':{'dice':3, 'focal':2, 'jaccard':3}})","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:10.488557Z","iopub.execute_input":"2021-12-24T03:09:10.488988Z","iopub.status.idle":"2021-12-24T03:09:10.627368Z","shell.execute_reply.started":"2021-12-24T03:09:10.488952Z","shell.execute_reply":"2021-12-24T03:09:10.626696Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# copy our file into the working directory (make sure it has .py suffix)\ncopyfile(src = \"../input/d/pauldark/u-net-loss/Unet_combo.pth\", dst = \"../working/Unet_combo.pth\")","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:10.628702Z","iopub.execute_input":"2021-12-24T03:09:10.628949Z","iopub.status.idle":"2021-12-24T03:09:10.930299Z","shell.execute_reply.started":"2021-12-24T03:09:10.628917Z","shell.execute_reply":"2021-12-24T03:09:10.929660Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"CHECKPOINT = 'Unet_combo.pth'","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:10.931888Z","iopub.execute_input":"2021-12-24T03:09:10.932623Z","iopub.status.idle":"2021-12-24T03:09:10.936481Z","shell.execute_reply.started":"2021-12-24T03:09:10.932586Z","shell.execute_reply":"2021-12-24T03:09:10.935712Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"if LOAD_MODEL:\n load_checkpoint(\n CHECKPOINT, model, optimizer, LR,\n )","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:10.944106Z","iopub.execute_input":"2021-12-24T03:09:10.944295Z","iopub.status.idle":"2021-12-24T03:09:11.067304Z","shell.execute_reply.started":"2021-12-24T03:09:10.944272Z","shell.execute_reply":"2021-12-24T03:09:11.066522Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# history = train(model, optimizer, criterion, train_loader, val_loader, model_name=CHECKPOINT)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:11.069477Z","iopub.execute_input":"2021-12-24T03:09:11.070107Z","iopub.status.idle":"2021-12-24T03:09:11.073957Z","shell.execute_reply.started":"2021-12-24T03:09:11.070069Z","shell.execute_reply":"2021-12-24T03:09:11.073096Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Let' download weights of trained model.","metadata":{}},{"cell_type":"code","source":"# Load the latest model data_dir\n# if LOAD_MODEL:\n# load_checkpoint(\n# CHECKPOINT, model, optimizer, LR,\n# )\n \n# metrics = eval_loop(model, criterion, val_loader)\n# print('accuracy:', metrics['accuracy'])\n# print('f1 macro:', metrics['f1_macro'])\n# print('test loss:', metrics['loss'])","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:11.075207Z","iopub.execute_input":"2021-12-24T03:09:11.075994Z","iopub.status.idle":"2021-12-24T03:09:11.082949Z","shell.execute_reply.started":"2021-12-24T03:09:11.075958Z","shell.execute_reply":"2021-12-24T03:09:11.082217Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# #del metrics\n# torch.cuda.empty_cache()\n# del model","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:11.084301Z","iopub.execute_input":"2021-12-24T03:09:11.084559Z","iopub.status.idle":"2021-12-24T03:09:11.090685Z","shell.execute_reply.started":"2021-12-24T03:09:11.084525Z","shell.execute_reply":"2021-12-24T03:09:11.089928Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

Visualize Training Results

","metadata":{}},{"cell_type":"markdown","source":"Let's check loss and accuracies during the training of the model","metadata":{}},{"cell_type":"code","source":"def plot_loss_acc(history):\n \"\"\"Print Loss in train and val sets\"\"\"\n train_loss, val_loss, val_acc, _ = zip(*history)\n\n fig, (ax1, ax2) = plt.subplots(2, figsize=(10,10))\n fig.suptitle('Loss and Accuracy')\n ax1.plot(train_loss, label=\"train_loss\")\n ax1.plot(val_loss, label=\"val_loss\")\n ax1.legend(loc='best')\n plt.ylabel(\"loss\")\n\n ax2.plot(val_acc, label=\"val_accuracy\")\n ax2.legend(loc='best')\n plt.xlabel(\"epochs\")\n plt.ylabel(\"accuracy\")\n plt.show()","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:11.091945Z","iopub.execute_input":"2021-12-24T03:09:11.092200Z","iopub.status.idle":"2021-12-24T03:09:11.099234Z","shell.execute_reply.started":"2021-12-24T03:09:11.092164Z","shell.execute_reply":"2021-12-24T03:09:11.098520Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# plot_loss_acc(history)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:11.100631Z","iopub.execute_input":"2021-12-24T03:09:11.101118Z","iopub.status.idle":"2021-12-24T03:09:11.107520Z","shell.execute_reply.started":"2021-12-24T03:09:11.101081Z","shell.execute_reply":"2021-12-24T03:09:11.106702Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Inspired by:\n\n[Ant 🐜 ](https://www.kaggle.com/superant/simple-nn-segmentation/notebook)","metadata":{}},{"cell_type":"code","source":"def get_threshold(Y, pred):\n '''Function is called for finding threshols and IoU\n '''\n scores = list(pred.ravel())\n mask = list(Y.ravel())\n \n #idx from the end to the begining\n idxs=np.argsort(scores)[::-1]\n mask_sorted=np.array(mask)[idxs]\n sum_mask_one=np.cumsum(mask_sorted)\n IoU=sum_mask_one/(np.arange(1,len(mask_sorted)+1)+np.sum(mask_sorted)-sum_mask_one)\n best_IoU_idx=IoU.argmax()\n best_threshold=scores[idxs[best_IoU_idx]]\n best_IoU=IoU[best_IoU_idx]\n\n return best_threshold, best_IoU","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:11.108939Z","iopub.execute_input":"2021-12-24T03:09:11.109315Z","iopub.status.idle":"2021-12-24T03:09:11.117720Z","shell.execute_reply.started":"2021-12-24T03:09:11.109279Z","shell.execute_reply":"2021-12-24T03:09:11.116965Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Re-assign val_loader with higher num of batches (for further plotting)\n","metadata":{}},{"cell_type":"code","source":"val_loader_plot = DataLoader(val_dataset,\n batch_size=16,\n shuffle=False,\n pin_memory=True,\n num_workers=WORKERS,)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:11.119071Z","iopub.execute_input":"2021-12-24T03:09:11.119408Z","iopub.status.idle":"2021-12-24T03:09:11.126219Z","shell.execute_reply.started":"2021-12-24T03:09:11.119328Z","shell.execute_reply":"2021-12-24T03:09:11.125417Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Define function to get the best value of threshold for the IoU metric","metadata":{}},{"cell_type":"code","source":"@torch.no_grad()\ndef get_best_threshold(model,loader,device=DEVICE,N=3):\n img_thresholds = []\n img_IoUs = []\n model.eval()\n \n for img, mask in tqdm(itertools.islice(loader, N), total=N):\n img = img.to(device)\n mask = mask.detach().numpy()\n \n Y_hat = torch.sigmoid(model(img)).cpu().detach().numpy()\n \n for i in range(mask.shape[0]):\n best_img_threshold, best_img_IoU = get_threshold(mask[i], Y_hat[i])\n img_thresholds.append(best_img_threshold)\n img_IoUs.append(best_img_IoU)\n del model #！！！！！！！！！！！！！！！！！！！ \n return img_thresholds,img_IoUs","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:11.127594Z","iopub.execute_input":"2021-12-24T03:09:11.127883Z","iopub.status.idle":"2021-12-24T03:09:11.135611Z","shell.execute_reply.started":"2021-12-24T03:09:11.127847Z","shell.execute_reply":"2021-12-24T03:09:11.134710Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"img_thresholds,img_IoUs = get_best_threshold(model,val_loader_plot)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:09:11.137146Z","iopub.execute_input":"2021-12-24T03:09:11.137564Z","iopub.status.idle":"2021-12-24T03:12:14.026543Z","shell.execute_reply.started":"2021-12-24T03:09:11.137529Z","shell.execute_reply":"2021-12-24T03:12:14.025227Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"best_threshold = np.mean(img_thresholds)\nthreshold_std = np.std(img_thresholds)\navg_IoU = np.mean(img_IoUs)\n\nprint(f\"Best threshold: {best_threshold:.3g} (STD: +-{threshold_std:.3g}), Avg. Val IoU: {avg_IoU:.3f}\")","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:12:14.028135Z","iopub.execute_input":"2021-12-24T03:12:14.028676Z","iopub.status.idle":"2021-12-24T03:12:14.035771Z","shell.execute_reply.started":"2021-12-24T03:12:14.028613Z","shell.execute_reply":"2021-12-24T03:12:14.034900Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"threshold = best_threshold\nmodel.eval()\n\nX, Y = next(iter(val_loader_plot))\nX = X.to(DEVICE)\nY = Y.detach().numpy()\n\nwith torch.no_grad():\n pred=torch.sigmoid(model(X)).cpu().detach().numpy()\n \npred_Y = (pred >= threshold)\ndel X","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:12:14.037021Z","iopub.execute_input":"2021-12-24T03:12:14.037427Z","iopub.status.idle":"2021-12-24T03:14:14.221993Z","shell.execute_reply.started":"2021-12-24T03:12:14.037390Z","shell.execute_reply":"2021-12-24T03:14:14.221070Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Define a function to plot Real mask, Predicted mask and Overlapping mask with following colors:\n\n***green: correct prediction***\n\n***gray: false positive (too much)***\n\n***orange: false negative (missed)***","metadata":{}},{"cell_type":"code","source":"def plot(Y_val, Y_hat):\n output = np.zeros_like(Y_val)\n output = np.where((Y_val == 0) & (Y_hat == 1), 1, output)\n output = np.where((Y_val == 1) & (Y_hat == 0), 2, output)\n output = np.where((Y_val == 1) & (Y_hat == 1), 3, output)\n\n \n fig, (ax1, ax2, ax3) = plt.subplots(1, 3, figsize=(20, 15),)\n ax1.imshow(Y_val,cmap='gray')\n ax2.imshow(((Y_hat * 255)), cmap='gray')\n ax3.imshow(output, cmap=cmap)\n ax1.set_axis_off()\n ax2.set_axis_off()\n ax3.set_axis_off()\n ax1.set_title('Real')\n ax2.set_title('Output')\n ax3.set_title('Real+Output')\n plt.xticks([])\n plt.yticks([]);","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:14.223738Z","iopub.execute_input":"2021-12-24T03:14:14.224013Z","iopub.status.idle":"2021-12-24T03:14:14.232117Z","shell.execute_reply.started":"2021-12-24T03:14:14.223973Z","shell.execute_reply":"2021-12-24T03:14:14.231106Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"N = 5\nfor i in range(N):\n img_Y = Y[i, 0]\n img_pred = pred_Y[i, 0]\n \n plot(img_Y, img_pred)\n plt.show()","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:14.233437Z","iopub.execute_input":"2021-12-24T03:14:14.233945Z","iopub.status.idle":"2021-12-24T03:14:15.490466Z","shell.execute_reply.started":"2021-12-24T03:14:14.233909Z","shell.execute_reply":"2021-12-24T03:14:15.489792Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

Save Results to a Pandas DataFrame

","metadata":{}},{"cell_type":"code","source":"def get_scores(report_df,\n name,loss_name, img_IoUs, \n img_thresholds,metrics = None,\n history = None,\n load_weight = False):\n '''Create and add metrics into a pandas DF after experiment'''\n \n report = pd.DataFrame(columns={'Criterion'}, data=[0])\n \n if load_weight:\n train_loss, val_loss, val_acc, f1_macro = zip(*history)\n report['Validation_loss'] = val_loss[-1]\n report['Val_Acc_at_last_epoch'] = val_acc[-1]\n report['Mean_acc_in_val_set'] = np.mean(val_acc)\n report['F1_macro'] = f1_macro[-1]\n report['mean_F1_along_epochs'] = np.mean(f1_macro)\n\n \n report['Criterion'] = loss_name\n report['Average_Val_IoU'] = np.mean(img_IoUs)\n report['Best_threshold'] = np.mean(img_thresholds)\n\n report.index = [name]\n report_df = report_df.append(report)\n return report_df","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:15.491768Z","iopub.execute_input":"2021-12-24T03:14:15.492141Z","iopub.status.idle":"2021-12-24T03:14:15.502486Z","shell.execute_reply.started":"2021-12-24T03:14:15.492100Z","shell.execute_reply":"2021-12-24T03:14:15.501711Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"df_report = pd.DataFrame(data=None)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:15.503871Z","iopub.execute_input":"2021-12-24T03:14:15.504142Z","iopub.status.idle":"2021-12-24T03:14:15.516372Z","shell.execute_reply.started":"2021-12-24T03:14:15.504103Z","shell.execute_reply":"2021-12-24T03:14:15.515570Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"df_report = get_scores(df_report, 'Unet_BaseLine', 'ComboLoss',\n img_IoUs,img_thresholds)\ndf_report","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:15.517965Z","iopub.execute_input":"2021-12-24T03:14:15.518266Z","iopub.status.idle":"2021-12-24T03:14:15.534736Z","shell.execute_reply.started":"2021-12-24T03:14:15.518230Z","shell.execute_reply":"2021-12-24T03:14:15.533951Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

SUBMISSION

","metadata":{}},{"cell_type":"markdown","source":"Define function for encoding masks","metadata":{}},{"cell_type":"code","source":"class NeuroDataSet(Dataset):\n def __init__(self, files, transform=None): \n self.files = files\n self.height = HEIGHT\n self.width = WIDTH\n self.transform = transform\n self.len_ = len(self.files)\n \n \n def load_sample(self, file):\n img = cv2.imread(str(file), cv2.IMREAD_COLOR)\n img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n img_id = file.stem\n return img, img_id\n \n def __getitem__(self,index):\n # load image\n img, img_id = self.load_sample(self.files[index])\n \n if self.transform is not None:\n augmented = self.transform(image = img)\n img = augmented['image']\n return img,img_id\n else:\n return img,img_id\n \n def __len__(self):\n return self.len_","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:15.535920Z","iopub.execute_input":"2021-12-24T03:14:15.536567Z","iopub.status.idle":"2021-12-24T03:14:15.547371Z","shell.execute_reply.started":"2021-12-24T03:14:15.536527Z","shell.execute_reply":"2021-12-24T03:14:15.546642Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"test_dataset = NeuroDataSet(test_files,\n transform=get_augmentation(224, 'test')\n )","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:15.548747Z","iopub.execute_input":"2021-12-24T03:14:15.549032Z","iopub.status.idle":"2021-12-24T03:14:15.556736Z","shell.execute_reply.started":"2021-12-24T03:14:15.548989Z","shell.execute_reply":"2021-12-24T03:14:15.555993Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"test_loader = DataLoader(test_dataset,\n batch_size=3,\n shuffle=False,\n pin_memory=True,\n num_workers=WORKERS,)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:15.558249Z","iopub.execute_input":"2021-12-24T03:14:15.558507Z","iopub.status.idle":"2021-12-24T03:14:15.568342Z","shell.execute_reply.started":"2021-12-24T03:14:15.558469Z","shell.execute_reply":"2021-12-24T03:14:15.567613Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Utilities","metadata":{}},{"cell_type":"code","source":"@torch.no_grad()\ndef predict_mask(image,model,device=DEVICE):\n model.eval()\n# image = image.view(-1,3,224,224)\n image = image.to(device)\n y_pred = torch.sigmoid(model(image)).cpu().detach().numpy()\n# y_pred = np.squeeze(y_pred,axis=0)\n return y_pred","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:15.569556Z","iopub.execute_input":"2021-12-24T03:14:15.570298Z","iopub.status.idle":"2021-12-24T03:14:15.578412Z","shell.execute_reply.started":"2021-12-24T03:14:15.570260Z","shell.execute_reply":"2021-12-24T03:14:15.577714Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"img, im_id = next(iter(test_loader))\npreds_test = predict_mask(img,model)\n#preds_test = preds_test >= threshold","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:15.579590Z","iopub.execute_input":"2021-12-24T03:14:15.580443Z","iopub.status.idle":"2021-12-24T03:14:16.744003Z","shell.execute_reply.started":"2021-12-24T03:14:15.580403Z","shell.execute_reply":"2021-12-24T03:14:16.743123Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def mask_convert(mask):\n #mask = mask.clone().cpu().detach().numpy()\n mask = mask.transpose((1,2,0))\n mask = np.squeeze(mask)\n return mask","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:16.747705Z","iopub.execute_input":"2021-12-24T03:14:16.747946Z","iopub.status.idle":"2021-12-24T03:14:16.753360Z","shell.execute_reply.started":"2021-12-24T03:14:16.747916Z","shell.execute_reply":"2021-12-24T03:14:16.751371Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def rle_encoding(x):\n dots = np.where(x.flatten() == 1)[0]\n run_lengths = []\n prev = -2\n for b in dots:\n if (b>prev+1): run_lengths.extend((b + 1, 0))\n run_lengths[-1] += 1\n prev = b\n return ' '.join(map(str, run_lengths))","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:16.754882Z","iopub.execute_input":"2021-12-24T03:14:16.755149Z","iopub.status.idle":"2021-12-24T03:14:16.763244Z","shell.execute_reply.started":"2021-12-24T03:14:16.755114Z","shell.execute_reply":"2021-12-24T03:14:16.762443Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"for pred in preds_test:\n pred = mask_convert(pred)\n plt.imshow(pred)\n plt.title('pred before resize')\n plt.axis(\"off\")\n plt.show()","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:16.764635Z","iopub.execute_input":"2021-12-24T03:14:16.764931Z","iopub.status.idle":"2021-12-24T03:14:17.080797Z","shell.execute_reply.started":"2021-12-24T03:14:16.764896Z","shell.execute_reply":"2021-12-24T03:14:17.080034Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"#test_mask: after reshape before fix_overlapping\ntest_masks = []\nfor pred in preds_test:\n pred = mask_convert(pred)\n test_mask=cv2.resize(pred,dsize=(704,520),interpolation=cv2.INTER_CUBIC).reshape(520,704,1)\n print(test_mask.shape)\n test_mask = test_mask >= threshold\n test_masks.append(test_mask)\n \n plt.imshow(test_mask)\n plt.title('pred after resize')\n plt.axis(\"off\")\n plt.show()","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:17.081998Z","iopub.execute_input":"2021-12-24T03:14:17.082762Z","iopub.status.idle":"2021-12-24T03:14:17.736439Z","shell.execute_reply.started":"2021-12-24T03:14:17.082708Z","shell.execute_reply":"2021-12-24T03:14:17.735547Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"print(test_masks[0].shape)\n","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:17.737601Z","iopub.execute_input":"2021-12-24T03:14:17.738072Z","iopub.status.idle":"2021-12-24T03:14:17.744032Z","shell.execute_reply.started":"2021-12-24T03:14:17.738006Z","shell.execute_reply":"2021-12-24T03:14:17.743176Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def check_overlap(msk):\n msk = msk.astype(np.bool).astype(np.uint8)\n return np.any(np.sum(msk, axis=-1)>1)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:17.746012Z","iopub.execute_input":"2021-12-24T03:14:17.746275Z","iopub.status.idle":"2021-12-24T03:14:17.754916Z","shell.execute_reply.started":"2021-12-24T03:14:17.746242Z","shell.execute_reply":"2021-12-24T03:14:17.753987Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import tensorflow as tf","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:17.756066Z","iopub.execute_input":"2021-12-24T03:14:17.756477Z","iopub.status.idle":"2021-12-24T03:14:19.035112Z","shell.execute_reply.started":"2021-12-24T03:14:17.756439Z","shell.execute_reply":"2021-12-24T03:14:19.034293Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def fix_overlap(msk):\n \"\"\"\n Args:\n mask: multi-channel mask, each channel is an instance of cell, shape:(520,704,None)\n Returns:\n multi-channel mask with non-overlapping values, shape:(520,704,None)\n \"\"\"\n msk = np.array(msk)\n msk = np.pad(msk, [[0,0],[0,0],[1,0]])\n ins_len = msk.shape[-1]\n msk = np.argmax(msk,axis=-1)\n msk = tf.keras.utils.to_categorical(msk, num_classes=ins_len)\n msk = msk[...,1:]\n msk = msk[...,np.any(msk, axis=(0,1))]\n return msk","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:19.036291Z","iopub.execute_input":"2021-12-24T03:14:19.038209Z","iopub.status.idle":"2021-12-24T03:14:19.047037Z","shell.execute_reply.started":"2021-12-24T03:14:19.038178Z","shell.execute_reply":"2021-12-24T03:14:19.046287Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"for test_mask in test_masks:\n overlap_test_masks=check_overlap(test_mask)\n print(overlap_test_masks)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:19.048448Z","iopub.execute_input":"2021-12-24T03:14:19.048760Z","iopub.status.idle":"2021-12-24T03:14:19.063108Z","shell.execute_reply.started":"2021-12-24T03:14:19.048712Z","shell.execute_reply":"2021-12-24T03:14:19.062312Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"#test_mask2: after reshape after fix_overlapping\ntest_masks2=[]\nfor test_mask in test_masks:\n test_mask2 = fix_overlap(test_mask).reshape(520,704,1)\n print(test_mask2.shape)\n test_masks2+=[test_mask2]","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:19.064503Z","iopub.execute_input":"2021-12-24T03:14:19.064784Z","iopub.status.idle":"2021-12-24T03:14:19.239413Z","shell.execute_reply.started":"2021-12-24T03:14:19.064748Z","shell.execute_reply":"2021-12-24T03:14:19.238702Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"for test_mask2 in test_masks2:\n overlap_test_masks2=check_overlap(test_mask2)\n print(overlap_test_masks2)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:19.240695Z","iopub.execute_input":"2021-12-24T03:14:19.241113Z","iopub.status.idle":"2021-12-24T03:14:19.248745Z","shell.execute_reply.started":"2021-12-24T03:14:19.241076Z","shell.execute_reply":"2021-12-24T03:14:19.248039Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"predicted2 = [rle_encoding(test_mask2) for test_mask2 in test_masks2]","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:19.250149Z","iopub.execute_input":"2021-12-24T03:14:19.250619Z","iopub.status.idle":"2021-12-24T03:14:19.345789Z","shell.execute_reply.started":"2021-12-24T03:14:19.250580Z","shell.execute_reply":"2021-12-24T03:14:19.345088Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"for img,img_name in test_loader:\n masks = predict_mask(img,model)\n print(masks.shape)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:19.347127Z","iopub.execute_input":"2021-12-24T03:14:19.347392Z","iopub.status.idle":"2021-12-24T03:14:20.000625Z","shell.execute_reply.started":"2021-12-24T03:14:19.347356Z","shell.execute_reply":"2021-12-24T03:14:19.999017Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"sample_submission=pd.read_csv(SAMPLE_SUBMISSION)\nsample_submission","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:20.002386Z","iopub.execute_input":"2021-12-24T03:14:20.002982Z","iopub.status.idle":"2021-12-24T03:14:20.019360Z","shell.execute_reply.started":"2021-12-24T03:14:20.002939Z","shell.execute_reply":"2021-12-24T03:14:20.018567Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"submit = sample_submission.copy()\nsubmit['predicted'] = predicted2\n#submit.to_csv('submission.csv', index=False)\nsubmit","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:20.020443Z","iopub.execute_input":"2021-12-24T03:14:20.020715Z","iopub.status.idle":"2021-12-24T03:14:20.030652Z","shell.execute_reply.started":"2021-12-24T03:14:20.020678Z","shell.execute_reply":"2021-12-24T03:14:20.029728Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"ids=sample_submission['id'].tolist()\nprint(ids)\nsubmit2=pd.DataFrame(columns=sample_submission.columns) \n\nfor idi, test_mask2 in zip(ids,test_masks2):\n annos=rle_encoding(test_mask2)\n annos2=annos.split(' ')\n annos4=[]\n for i in range(len(annos2)//100):\n annos3=''\n for j in range(100):\n annos3+=annos2[i*100+j]+' '\n annos4+=[annos3]\n \n submit=pd.DataFrame(columns=sample_submission.columns) \n submit['predicted']=annos4\n submit['id']=idi\n submit2=pd.concat([submit2,submit],axis=0)\nsubmit2=submit2.reset_index(drop=True)\n \nsubmit2.to_csv('submission.csv', index=False)\nsubmit2","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:20.032383Z","iopub.execute_input":"2021-12-24T03:14:20.032848Z","iopub.status.idle":"2021-12-24T03:14:20.158685Z","shell.execute_reply.started":"2021-12-24T03:14:20.032812Z","shell.execute_reply":"2021-12-24T03:14:20.157886Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"plt.imshow(test_masks[0])\nplt.title('before fix_overlapping')\nplt.axis(\"off\")\nplt.show()\n\nplt.imshow(test_masks2[0])\nplt.title('after fix_overlapping')\nplt.axis(\"off\")\nplt.show()","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:20.159905Z","iopub.execute_input":"2021-12-24T03:14:20.160210Z","iopub.status.idle":"2021-12-24T03:14:20.436944Z","shell.execute_reply.started":"2021-12-24T03:14:20.160170Z","shell.execute_reply":"2021-12-24T03:14:20.436251Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## NERABOTAYUSHAYA HERNYA","metadata":{}},{"cell_type":"code","source":"# def rle_encoding(x):\n# dots = np.where(x.flatten() == 1)[0]\n# run_lengths = []\n# prev = -2\n# for b in dots:\n# if (b>prev+1): run_lengths.extend((b + 1, 0))\n# run_lengths[-1] += 1\n# prev = b\n# return ' '.join(map(str, run_lengths))\n\n\n# def remove_overlapping_pixels(mask, other_masks):\n# for other_mask in other_masks:\n# if np.sum(np.logical_and(mask, other_mask)) > 0:\n# mask[np.logical_and(mask, other_mask)] = 0\n# return mask","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:20.438274Z","iopub.execute_input":"2021-12-24T03:14:20.439021Z","iopub.status.idle":"2021-12-24T03:14:20.443347Z","shell.execute_reply.started":"2021-12-24T03:14:20.438986Z","shell.execute_reply":"2021-12-24T03:14:20.442384Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# # converting tensor to image\n# def image_convert(image):\n# image = image.clone().cpu().numpy()\n# image = image.transpose((1,2,0))\n# mean = np.array([0.485, 0.456, 0.406])\n# std = np.array([0.229, 0.224, 0.225])\n# image = std * image + mean\n# image = image.clip(0,1)\n# image = (image * 255).astype(np.uint8)\n# return image\n\n# def mask_convert(mask):\n# #mask = mask.clone().cpu().detach().numpy()\n# mask = mask.transpose((1,2,0))\n# mask = np.squeeze(mask)\n# return mask","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:20.444620Z","iopub.execute_input":"2021-12-24T03:14:20.445322Z","iopub.status.idle":"2021-12-24T03:14:20.452276Z","shell.execute_reply.started":"2021-12-24T03:14:20.445285Z","shell.execute_reply":"2021-12-24T03:14:20.451325Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# def plot(img, Y_hat):\n \n# img = image_convert(img)\n# Y_hat = mask_convert(Y_hat)\n# fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(20, 15),)\n# ax1.imshow(img,)\n# ax2.imshow((Y_hat), cmap='seismic')\n# # ax2.imshow(((Y_hat * 255)), cmap='gray')\n# ax1.set_axis_off()\n# ax2.set_axis_off()\n# ax1.set_title('Real')\n# ax2.set_title('Output')\n# plt.xticks([])\n# plt.yticks([]);","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:20.453616Z","iopub.execute_input":"2021-12-24T03:14:20.453988Z","iopub.status.idle":"2021-12-24T03:14:20.463100Z","shell.execute_reply.started":"2021-12-24T03:14:20.453955Z","shell.execute_reply":"2021-12-24T03:14:20.462382Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# @torch.no_grad()\n# def predict_mask(image,model,device=DEVICE):\n# model.eval()\n# image = image.view(-1,3,224,224)\n# image = image.to(device)\n# y_pred = torch.sigmoid(model(image)).cpu().detach().numpy()\n# y_pred = np.squeeze(y_pred,axis=0)\n# return y_pred","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:20.466035Z","iopub.execute_input":"2021-12-24T03:14:20.466241Z","iopub.status.idle":"2021-12-24T03:14:20.471394Z","shell.execute_reply.started":"2021-12-24T03:14:20.466217Z","shell.execute_reply":"2021-12-24T03:14:20.470678Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# threshold = best_threshold\n# submission = []\n# img, img_id = next(iter(test_loader))\n# for i, im_id in enumerate(img_id):\n# previous_masks = []\n# result = predict_mask(img[i],model) \n# binary_mask = (result >= threshold) \n# #binary_mask = result > 0.5\n# binary_mask = remove_overlapping_pixels(binary_mask, previous_masks)\n# previous_masks.append(binary_mask)\n# rle = rle_encoding(binary_mask)\n# submission.append((im_id, rle))\n# plot(img[i],binary_mask)\n \n# all_images_ids = [im_id for im_id, rle in submission]\n# if im_id not in all_images_ids:\n# submission.append((im_id, \"\"))\n","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:20.474448Z","iopub.execute_input":"2021-12-24T03:14:20.474676Z","iopub.status.idle":"2021-12-24T03:14:20.480905Z","shell.execute_reply.started":"2021-12-24T03:14:20.474627Z","shell.execute_reply":"2021-12-24T03:14:20.480136Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# df_sub = pd.DataFrame(submission, columns=['id', 'predicted'])\n# df_sub.to_csv(\"submission.csv\", index=False)\n# df_sub","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:20.482103Z","iopub.execute_input":"2021-12-24T03:14:20.482494Z","iopub.status.idle":"2021-12-24T03:14:20.492920Z","shell.execute_reply.started":"2021-12-24T03:14:20.482388Z","shell.execute_reply":"2021-12-24T03:14:20.492185Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"HEROVAYA HERNYA","metadata":{}},{"cell_type":"code","source":"# test_dataset = NeuroDataSet(test_files,\n# transform=get_augmentation(224, 'test')\n# )\n\n# test_loader = DataLoader(test_dataset,\n# batch_size=3,\n# shuffle=False,\n# pin_memory=True,\n# num_workers=WORKERS,)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:20.494272Z","iopub.execute_input":"2021-12-24T03:14:20.494668Z","iopub.status.idle":"2021-12-24T03:14:20.500990Z","shell.execute_reply.started":"2021-12-24T03:14:20.494592Z","shell.execute_reply":"2021-12-24T03:14:20.500357Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# for image,img_id in test_loader:\n# print(img_id)\n# print(image.shape)\n# break","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:20.502405Z","iopub.execute_input":"2021-12-24T03:14:20.502753Z","iopub.status.idle":"2021-12-24T03:14:20.509065Z","shell.execute_reply.started":"2021-12-24T03:14:20.502631Z","shell.execute_reply":"2021-12-24T03:14:20.508267Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# @torch.no_grad()\n# def predict_mask(image,model,device=DEVICE):\n# model.eval()\n# image = image.to(device)\n# y_pred = torch.sigmoid(model(image)).cpu().detach().numpy()\n# return y_pred\n\n# def mask_convert(mask):\n# #mask = mask.clone().cpu().detach().numpy()\n# mask = mask.transpose((1,2,0))\n# mask = np.squeeze(mask)\n# return mask\n\n# # converting tensor to image\n# def image_convert(image):\n# image = image.clone().cpu().numpy()\n# image = image.transpose((1,2,0))\n# mean = np.array([0.485, 0.456, 0.406])\n# std = np.array([0.229, 0.224, 0.225])\n# image = std * image + mean\n# image = image.clip(0,1)\n# image = (image * 255).astype(np.uint8)\n# return image","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:20.510467Z","iopub.execute_input":"2021-12-24T03:14:20.510770Z","iopub.status.idle":"2021-12-24T03:14:20.517630Z","shell.execute_reply.started":"2021-12-24T03:14:20.510722Z","shell.execute_reply":"2021-12-24T03:14:20.516844Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# for img,img_names in test_loader:\n# masks = predict_mask(img,model)\n# for i,mask in enumerate(masks):\n# mask = mask_convert(mask)\n# mask = np.where(mask > 0.5, 1, 0)\n# image = img[i]\n# image = image_convert(image)\n# fig,axes = plt.subplots(1,2,figsize=(16,7))\n# axes[0].imshow(image)\n# axes[0].set_title('Image')\n# axes[1].imshow(mask,cmap=\"gray\")\n# axes[1].set_title('Predicted Mask')\n# plt.suptitle(img_names[i])\n# plt.tight_layout()\n# plt.show()","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:20.519367Z","iopub.execute_input":"2021-12-24T03:14:20.519747Z","iopub.status.idle":"2021-12-24T03:14:20.526124Z","shell.execute_reply.started":"2021-12-24T03:14:20.519704Z","shell.execute_reply":"2021-12-24T03:14:20.525468Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# def ins2rle(ins):\n# '''\n# img: numpy array, 1 - mask, 0 - background\n# Returns run length as string formated\n# '''\n# ins = np.array(ins)\n# pixels = ins.flatten()\n# pad = np.array([0])\n# pixels = np.concatenate([pad, pixels, pad])\n# runs = np.where(pixels[1:] != pixels[:-1])[0] + 1\n# runs[1::2] -= runs[::2]\n# rles = ' '.join(str(x) for x in runs)\n# return rles\n\n# def mask2rle(mask, cutoff=0.5, min_object_size=1.0):\n# \"\"\" Return run length encoding of mask. \n# ref: https://www.kaggle.com/raoulma/nuclei-dsb-2018-tensorflow-u-net-score-0-352\n# \"\"\"\n# # segment image and label different objects\n# lab_mask = skimage.morphology.label(mask > cutoff)\n \n \n# # Keep only objects that are large enough.\n# (mask_labels, mask_sizes) = np.unique(lab_mask, return_counts=True)\n \n\n# if (mask_sizes < min_object_size).any():\n# # print(mask_sizes)\n# mask_labels = mask_labels[mask_sizes < min_object_size]\n \n# for n in mask_labels:\n# lab_mask[lab_mask == n] = 0\n# lab_mask = skimage.morphology.label(lab_mask > cutoff) \n \n# rles = []\n# # Loop over each object excluding the background labeled by 0.\n# for i in range(1, lab_mask.max() + 1):\n# rle = ins2rle(lab_mask == i)\n# rles.append(rle)\n \n# return rles\n","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:20.527555Z","iopub.execute_input":"2021-12-24T03:14:20.527857Z","iopub.status.idle":"2021-12-24T03:14:20.537182Z","shell.execute_reply.started":"2021-12-24T03:14:20.527825Z","shell.execute_reply":"2021-12-24T03:14:20.536464Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# pred_rles , pred_paths = [],[]\n# for img,img_name in test_loader:\n# masks = predict_mask(img,model)\n# for i,mask in enumerate(masks):\n# mask = mask_convert(mask)\n# msk = cv2.resize(mask,(704,520),interpolation=cv2.INTER_NEAREST)\n# rle_list = mask2rle(msk)\n# img_id = img_name[i].split('.')[0]\n# idx_list = [img_id for i in range(len(rle_list))] \n# pred_rles.extend(rle_list)\n# pred_paths.extend(idx_list)","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:20.538411Z","iopub.execute_input":"2021-12-24T03:14:20.538806Z","iopub.status.idle":"2021-12-24T03:14:20.546140Z","shell.execute_reply.started":"2021-12-24T03:14:20.538770Z","shell.execute_reply":"2021-12-24T03:14:20.545421Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# df = pd.DataFrame({'id':pred_paths, 'predicted':pred_rles})\n# df.to_csv('submission.csv', index=False)\n# df = pd.read_csv('submission.csv')\n# df","metadata":{"execution":{"iopub.status.busy":"2021-12-24T03:14:20.547257Z","iopub.execute_input":"2021-12-24T03:14:20.548027Z","iopub.status.idle":"2021-12-24T03:14:20.555669Z","shell.execute_reply.started":"2021-12-24T03:14:20.547988Z","shell.execute_reply":"2021-12-24T03:14:20.555082Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]}]}