2020-12-03 22:32:21 +00:00
|
|
|
import torch
|
|
|
|
from torch.utils.data import Dataset
|
|
|
|
import torchvision.transforms as T
|
|
|
|
from torchvision import datasets
|
|
|
|
|
|
|
|
# Wrapper for basic pytorch datasets which re-wraps them into a format usable by ExtensibleTrainer.
|
|
|
|
class TorchDataset(Dataset):
|
|
|
|
def __init__(self, opt):
|
|
|
|
DATASET_MAP = {
|
|
|
|
"mnist": datasets.MNIST,
|
|
|
|
"fmnist": datasets.FashionMNIST,
|
|
|
|
"cifar10": datasets.CIFAR10,
|
2021-01-12 03:09:16 +00:00
|
|
|
"imagenet": datasets.ImageNet,
|
|
|
|
"imagefolder": datasets.ImageFolder
|
2020-12-03 22:32:21 +00:00
|
|
|
}
|
2021-01-12 03:09:16 +00:00
|
|
|
normalize = T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
|
|
|
|
if opt['train']:
|
|
|
|
transforms = [
|
|
|
|
T.RandomResizedCrop(opt['image_size']),
|
|
|
|
T.RandomHorizontalFlip(),
|
|
|
|
T.ToTensor(),
|
|
|
|
normalize,
|
|
|
|
]
|
|
|
|
else:
|
|
|
|
transforms = [
|
|
|
|
T.Resize(opt['val_resize']),
|
|
|
|
T.CenterCrop(opt['image_size']),
|
|
|
|
T.ToTensor(),
|
|
|
|
normalize,
|
|
|
|
]
|
2020-12-03 22:32:21 +00:00
|
|
|
transforms = T.Compose(transforms)
|
2021-01-12 03:09:16 +00:00
|
|
|
self.dataset = DATASET_MAP[opt['dataset']](transform=transforms, **opt['kwargs'])
|
2020-12-03 22:32:21 +00:00
|
|
|
self.len = opt['fixed_len'] if 'fixed_len' in opt.keys() else len(self.dataset)
|
|
|
|
|
|
|
|
def __getitem__(self, item):
|
2021-01-12 03:09:16 +00:00
|
|
|
underlying_item, lbl = self.dataset[item]
|
|
|
|
return {'lq': underlying_item, 'hq': underlying_item, 'labels': lbl,
|
2020-12-03 22:32:21 +00:00
|
|
|
'LQ_path': str(item), 'GT_path': str(item)}
|
|
|
|
|
|
|
|
def __len__(self):
|
|
|
|
return self.len
|
2021-01-12 03:09:16 +00:00
|
|
|
|
|
|
|
if __name__ == '__main__':
|
|
|
|
opt = {
|
|
|
|
'flip': True,
|
|
|
|
'crop_sz': None,
|
|
|
|
'dataset': 'imagefolder',
|
|
|
|
'resize': 256,
|
|
|
|
'center_crop': 224,
|
|
|
|
'normalize': True,
|
|
|
|
'kwargs': {
|
|
|
|
'root': 'F:\\4k6k\\datasets\\images\\imagenet_2017\\val',
|
|
|
|
}
|
|
|
|
}
|
|
|
|
set = TorchDataset(opt)
|
|
|
|
j = set[0]
|