''' Split dataset as train, test, val 6:2:2 use function dataset_wrap, return {train:, val:, test:} torch dataset datasets names: isic2018, PH2, DMF, SKD ''' import os import json import torch import random import numpy as np from torchvision import transforms import albumentations as A import pandas as pd import pandas as pd dataset_indices = { 'isic2018': 0, 'PH2': 1, 'DMF': 2, 'SKD': 3, } def norm01(x): return np.clip(x, 0, 255) / 255 def Dataset_wrap_csv(k_fold='No', use_old_split=True, img_size=384, dataset_name='isic2018', split_ratio=[0.8, 0.2], train_aug=False, data_folder='/bigdata/siyiplace/data/skin_lesion'): ''' use train val test csv to load the whole datasets in order to include domain (dataset) label if k_fold is a number, means we use k-fold to do experiments, load k_fold index data. default 5 folders if use_old_split, load existing train, test paths dataset_name: choose which dataset to load random split train val test set by split_ratio save train test id return train val test in a dic ''' data_dic = {} data_path = '{}/{}/'.format(data_folder, dataset_name) # do k fold loading if k_fold != 'No': if use_old_split: try: train_df = pd.read_csv(data_path+'train_meta_kfold_{}.csv'.format(k_fold), dtype={'ID': str}) test_df = pd.read_csv(data_path+'test_meta_kfold_{}.csv'.format(k_fold), dtype={'ID': str}) data_dic['train'] = SkinDataset_csv(dataset_name, img_size, train_df, use_aug=train_aug, data_path=data_path) data_dic['test'] = SkinDataset_csv(dataset_name, img_size, test_df, use_aug=False, data_path=data_path) data_size = len(data_dic['train'])+len(data_dic['test']) print('{} has {} samples, {} are used to train, {} are used to test. \n 5 Folder -- Use {}' .format(dataset_name, data_size, len(data_dic['train']), len(data_dic['test']), k_fold)) return data_dic except: print('No existing k_folder files, start creating new splitting....') print('use new split') df = pd.read_csv(data_path+'meta_{}.csv'.format(dataset_name),dtype={'ID': str}) data_size = len(df) # # random split train test based on train_ratio index_list = list(range(data_size)) random.Random(42).shuffle(index_list) split_size = int(data_size/5.0+0.5) # one split size, 5 splits split_ids = [0,split_size,split_size*2,split_size*3,split_size*4,len(index_list)] for i in range(5): train_df = df.iloc[index_list[:split_ids[i]]+index_list[split_ids[i+1]:]] test_df = df.iloc[index_list[split_ids[i]:split_ids[i+1]]] # save train, test csv train_df.to_csv(data_path+'train_meta_kfold_{}.csv'.format(i), header=df.columns, index=False) test_df.to_csv(data_path+'test_meta_kfold_{}.csv'.format(i), header=df.columns, index=False) train_df = pd.read_csv(data_path+'train_meta_kfold_{}.csv'.format(k_fold), dtype={'ID': str}) test_df = pd.read_csv(data_path+'test_meta_kfold_{}.csv'.format(k_fold), dtype={'ID': str}) data_dic['train'] = SkinDataset_csv(dataset_name, img_size, train_df, use_aug=train_aug, data_path=data_path) data_dic['test'] = SkinDataset_csv(dataset_name, img_size, test_df, use_aug=False, data_path=data_path) assert data_size == len(data_dic['train'])+len(data_dic['test']) print('Finish creating new 5 folders. {} has {} samples, {} are used to train, {} are used to test. \n 5 Folder -- Use {}' .format(dataset_name, data_size, len(train_df), len(test_df), k_fold)) return data_dic # do no k fold if use_old_split: # in case these files are not exist try: train_df = pd.read_csv(data_path+'train_meta_{}.csv'.format(int(split_ratio[0]*100)), dtype={'ID': str}) test_df = pd.read_csv(data_path+'test_meta_{}.csv'.format(int(split_ratio[1]*100)), dtype={'ID': str}) data_dic['train'] = SkinDataset_csv(dataset_name, img_size, train_df, use_aug=train_aug, data_path=data_path) data_dic['test'] = SkinDataset_csv(dataset_name, img_size, test_df, use_aug=False, data_path=data_path) data_size = len(data_dic['train'])+len(data_dic['test']) print('{} has {} samples, {} are used to train, {} are used to test. \n The split ratio is {}' .format(dataset_name, data_size, len(data_dic['train']), len(data_dic['test']), split_ratio)) return data_dic except: print('No existing split files, start creating new splitting....') print('use new split') df = pd.read_csv(data_path+'meta_{}.csv'.format(dataset_name),dtype={'ID': str}) data_size = len(df) # # random split train test based on train_ratio index_list = list(range(data_size)) random.Random(42).shuffle(index_list) train_df = df.iloc[index_list[: int(data_size*split_ratio[0])]] test_df = df.iloc[index_list[int(data_size*split_ratio[0]) : ]] print('{} has {} samples, {} are used to train, {} are used to test. \n The split ratio is {}' .format(dataset_name, data_size, len(train_df), len(test_df), split_ratio)) # save train, test csv train_df.to_csv(data_path+'train_meta_{}.csv'.format(int(split_ratio[0]*100)), header=df.columns, index=False) test_df.to_csv(data_path+'test_meta_{}.csv'.format(int(split_ratio[1]*100)), header=df.columns, index=False) data_dic['train'] = SkinDataset_csv(dataset_name, img_size, train_df, use_aug=train_aug, data_path=data_path) data_dic['test'] = SkinDataset_csv(dataset_name, img_size, test_df, use_aug=False, data_path=data_path) return data_dic class SkinDataset_csv(torch.utils.data.Dataset): def __init__(self, dataset_name, img_size, df, use_aug=False, data_path='/bigdata/siyiplace/data/skin_lesion/isic2018/'): super(SkinDataset_csv, self).__init__() self.dataset_name = dataset_name self.root_dir = data_path self.df = df self.use_aug = use_aug self.num_samples = len(self.df) p = 0.5 self.aug_transf = A.Compose([ A.Resize(img_size, img_size), A.GaussNoise(p=p), A.HorizontalFlip(p=p), A.VerticalFlip(p=p), A.ShiftScaleRotate(p=p), A.RandomBrightnessContrast(p=p), ]) self.transf = A.Compose([ A.Resize(img_size, img_size), ]) self.normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) def __getitem__(self, index): if torch.is_tensor(index): index = index.tolist() row = self.df.loc[self.df.index[index]] img_path = self.root_dir + 'Image/{}.npy'.format(row['ID']) label_path = self.root_dir + 'Label/{}.npy'.format(row['ID']) diagnosis = row['diagnosis'] diagnosis_id = row['diagnosis_id'] img_data = np.load(img_path) label_data = np.load(label_path) > 0.5 if self.use_aug: tsf = self.aug_transf(image=img_data.astype('uint8'), mask=label_data.astype('uint8')) else: tsf = self.transf(image=img_data.astype('uint8'), mask=label_data.astype('uint8')) img_data, label_data = tsf['image'], tsf['mask'] img_data = norm01(img_data) label_data = np.expand_dims(label_data, 0) img_data = torch.from_numpy(img_data).float() label_data = torch.from_numpy(label_data).float() img_data = img_data.permute(2, 0, 1) img_data = self.normalize(img_data) return{ 'ID': row['ID'], 'set_name': self.dataset_name, 'set_id': dataset_indices[self.dataset_name], 'image_path': img_path, 'label_path': label_path, 'diagnosis': diagnosis, 'diagnosis_id': diagnosis_id, # i.e., class label 'image': img_data, 'label': label_data, } def __len__(self): return self.num_samples # ================================================================================================================================== # Load the whole dataset and don't split train test sets class SkinClasDataset(torch.utils.data.Dataset): ''' Use csv file to load the whole dataset. Have diagnosis labels used for generate tsne ''' def __init__(self, dataset_name, img_size, data_folder='/bigdata/siyiplace/data/skin_lesion'): super(SkinClasDataset, self).__init__() self.dataset_name = dataset_name self.root_dir = '{}/{}/'.format(data_folder, dataset_name) self.df = pd.read_csv(self.root_dir+'meta_{}.csv'.format(dataset_name), dtype={'ID': str}) self.num_samples = len(self.df) self.transf = A.Compose([ A.Resize(img_size, img_size), ]) self.normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) def __getitem__(self, index): img_path = self.root_dir + 'Image/{}.npy'.format(self.df.loc[self.df.index[index], 'ID']) label_path = self.root_dir + 'Label/{}.npy'.format(self.df.loc[self.df.index[index], 'ID']) diagnosis = self.df.loc[self.df.index[index], 'diagnosis'] diagnosis_id = self.df.loc[self.df.index[index], 'diagnosis_id'] img_data = np.load(img_path) label_data = np.load(label_path) > 0.5 tsf = self.transf(image=img_data.astype('uint8'), mask=label_data.astype('uint8')) img_data, label_data = tsf['image'], tsf['mask'] img_data = norm01(img_data) label_data = np.expand_dims(label_data, 0) img_data = torch.from_numpy(img_data).float() label_data = torch.from_numpy(label_data).float() img_data = img_data.permute(2, 0, 1) img_data = self.normalize(img_data) return{ 'set_name': self.dataset_name, 'set_id': dataset_indices[self.dataset_name], 'image_path': img_path, 'label_path': label_path, 'diagnosis': diagnosis, 'diagnosis_id': diagnosis_id, 'image': img_data, 'label': label_data, } def __len__(self): return self.num_samples if __name__ == '__main__': # test datasets = Dataset_wrap_csv(k_fold='No', use_old_split=False, dataset_name='SKD',dynamic=True) for key in datasets.keys(): print(len(datasets[key])) dataloader = torch.utils.data.DataLoader(datasets['train'], batch_size=3, shuffle=True, num_workers=2, pin_memory=True, drop_last=True) batch = next(iter(dataloader)) print(batch['four_id']) pass