| """ |
| Author: Mélanie Gaillochet |
| Date: 2020-10-05 |
| """ |
| import os |
| import numpy as np |
| import cv2 |
| import matplotlib.pyplot as plt |
|
|
| import h5py |
| import torch |
| from torch.utils.data import Dataset |
|
|
| from Utils.utils import random_selection |
| from Utils.augmentation_utils import random_augmentation |
| from Utils.load_utils import load_single_image |
| from Utils.utils import natural_keys |
|
|
|
|
| class MyDataset(Dataset): |
| def __init__(self, data_folder, config): |
| self.data_folder = data_folder |
| self.dataset_name = config['dataset_name'] |
| self.num_items = config['num_items'] |
| try: |
| self.downsample_size = eval(config['downsample_size']) |
| except TypeError: |
| self.downsample_size = None |
| |
| self.training = False |
| self.augment = config['augment'] |
| if self.augment: |
| self.augmentations = eval(config['augmentations']) |
| print('\nAugmenting the data with {} \n'.format(self.augmentations)) |
| self.aug_gaussian_mean = config['aug_gaussian_mean'] if 'gaussian_noise' in self.augmentations else 0 |
| self.aug_gaussian_std = config['aug_gaussian_std'] if 'gaussian_noise' in self.augmentations else 0 |
| print('self.aug_gaussian_mean {}'.format(self.aug_gaussian_mean)) |
| print('self.aug_gaussian_std {}'.format(self.aug_gaussian_std)) |
|
|
| |
| self.volume_folder_path = os.path.join(self.data_folder, self.dataset_name, 'data') |
| with h5py.File(self.volume_folder_path + '.hdf5', 'r') as hf: |
| self.volume_list = list(hf.keys()) |
| self.volume_list.sort(key=natural_keys) |
|
|
| self.seg_folder_path = os.path.join(self.data_folder, self.dataset_name, 'label') |
| with h5py.File(self.seg_folder_path + '.hdf5', 'r') as hf: |
| self.seg_list = list(hf.keys()) |
| self.seg_list.sort(key=natural_keys) |
|
|
| |
| if self.num_items != "all": |
| self.volume_list, self.seg_list = random_selection(self.num_items, self.volume_list, |
| self.seg_list) |
|
|
| def __len__(self): |
| """We return the total number of samples""" |
| return len(self.volume_list) |
|
|
| def __getitem__(self, idx): |
| """We generate one sample of data""" |
| |
| img = load_single_image(self.volume_folder_path, self.volume_list, idx) |
| label = load_single_image(self.seg_folder_path, self.seg_list, idx) |
|
|
| volume = img.copy() |
| target = label.copy() |
|
|
| |
| if self.downsample_size is not None and volume.shape[-1] != self.downsample_size[-1]: |
| resized_vol = np.zeros((volume.shape[0],) + self.downsample_size) |
| for i in range(0, volume.shape[0]): |
| resized_vol[i, :, :] = cv2.resize(np.float32(volume[i, :, :]), |
| dsize=self.downsample_size, |
| interpolation=cv2.INTER_CUBIC) |
| volume = resized_vol.copy() |
| target = cv2.resize(np.float32(target), |
| dsize=self.downsample_size, |
| interpolation=cv2.INTER_NEAREST) |
|
|
| |
| volume = torch.from_numpy(volume) |
| target = torch.from_numpy(target.astype(float)).long() |
|
|
| if self.training and self.augment: |
| if len(img.shape) == 3: |
| flip_axis, rotaxis0, rotaxis1 = 1, 1, 2 |
| elif len(img.shape) == 4: |
| flip_axis, rotaxis0, rotaxis1 = 2, 2, 3 |
| |
| volume, aug_dic = random_augmentation(volume, None, flip_axis, rotaxis0, rotaxis1, self.augmentations, |
| aug_gaussian_mean=self.aug_gaussian_mean, aug_gaussian_std=self.aug_gaussian_std) |
| _target, _ = random_augmentation(target.unsqueeze(0), aug_dic, flip_axis, rotaxis0, rotaxis1, self.augmentations, type='target') |
| target = _target[0, :, :] |
| |
| return volume, target, idx |
|
|
