def ImageNet(batch_size=12, output_size=224, cache_dir='tmp') -> DataBundle:
size = output_size
__imagenet_pca = {
'eigval':
torch.Tensor([0.2175, 0.0188, 0.0045]),
'eigvec':
torch.Tensor([
[-0.5675, 0.7192, 0.4009],
[-0.5808, -0.0045, -0.8140],
[-0.5836, -0.6948, 0.4203],
])
}
train_tfms = transforms.Compose([
transforms.Resize(output_size),
transforms.CenterCrop(output_size),
transforms.RandomHorizontalFlip(),
transforms.ColorJitter(.4, .4, .4),
transforms.ToTensor(),
Lighting(0.1, __imagenet_pca['eigval'], __imagenet_pca['eigvec']),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
val_tfms = transforms.Compose([
transforms.Resize(int(output_size * 1.14)),
transforms.CenterCrop(output_size),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
trainset = torchvision.datasets.ImageNet(root="D:\\",
transform=train_tfms,
split='train')
testset = torchvision.datasets.ImageNet(root="D:\\",
transform=val_tfms,
split='val')
train_loader = torch.utils.data.DataLoader(trainset,
batch_size=batch_size,
shuffle=True,
num_workers=6,
pin_memory=False)
test_loader = torch.utils.data.DataLoader(testset,
batch_size=batch_size,
shuffle=False,
num_workers=6,
pin_memory=False)
train_loader.name = "ImageNet"
return DataBundle(dataset_name="ImageNet",
train_dataset=train_loader,
test_dataset=test_loader,
cardinality=1000,
output_resolution=output_size,
is_classifier=True)
def ReflectBorderCifar10(batch_size=12,
output_size=32,
cache_dir='tmp') -> DataBundle:
border = (output_size - 32) // 2
uneven = output_size % 2
# Transformations
RC = transforms.RandomCrop((32, 32), padding=4)
PAD = transforms.Pad((border, border, border + uneven, border + uneven),
padding_mode='reflect')
RHF = transforms.RandomHorizontalFlip()
NRM = transforms.Normalize((0.4914, 0.4822, 0.4465),
(0.2023, 0.1994, 0.2010))
TT = transforms.ToTensor()
RS = transforms.Resize(output_size)
# Transforms object for trainset with augmentation
transform_with_aug = transforms.Compose([RC, RHF, PAD, TT, NRM])
# Transforms object for testset with NO augmentation
transform_no_aug = transforms.Compose([PAD, TT, NRM])
trainset = torchvision.datasets.CIFAR10(root=cache_dir,
train=True,
download=True,
transform=transform_with_aug)
train_loader = torch.utils.data.DataLoader(trainset,
batch_size=batch_size,
shuffle=True,
num_workers=8,
pin_memory=True)
testset = torchvision.datasets.CIFAR10(root=cache_dir,
train=False,
download=True,
transform=transform_no_aug)
test_loader = torch.utils.data.DataLoader(testset,
batch_size=batch_size,
shuffle=False,
num_workers=8,
pin_memory=True)
train_loader.name = "ReflectBorderCifar10"
return DataBundle(dataset_name="ReflectBorderCifar10",
train_dataset=train_loader,
test_dataset=test_loader,
cardinality=10,
output_resolution=output_size,
is_classifier=True)
def Cifar100(batch_size=12,
output_size=(32, 32),
cache_dir='tmp') -> DataBundle:
# Transformations
RC = transforms.RandomCrop((32, 32), padding=4)
RHF = transforms.RandomHorizontalFlip()
NRM = transforms.Normalize((0.4914, 0.4822, 0.4465),
(0.2023, 0.1994, 0.2010))
TT = transforms.ToTensor()
RS = transforms.Resize(output_size)
# Transforms object for trainset with augmentation
transform_with_aug = transforms.Compose([RC, RHF, RS, TT, NRM])
# Transforms object for testset with NO augmentation
transform_no_aug = transforms.Compose([RS, TT, NRM])
trainset = torchvision.datasets.CIFAR100(root=cache_dir,
train=True,
download=True,
transform=transform_with_aug)
train_loader = torch.utils.data.DataLoader(trainset,
batch_size=batch_size,
shuffle=True,
num_workers=8)
testset = torchvision.datasets.CIFAR100(root=cache_dir,
train=False,
download=True,
transform=transform_no_aug)
test_loader = torch.utils.data.DataLoader(testset,
batch_size=batch_size,
shuffle=False,
num_workers=8)
train_loader.name = "Cifar100"
return DataBundle(dataset_name="Cifar100",
train_dataset=train_loader,
test_dataset=test_loader,
cardinality=100,
output_resolution=output_size,
is_classifier=True)
def TinyResizedFood101(batch_size=12,
output_size=256,
cache_dir='tmp',
shuffle_test: bool = False) -> DataBundle:
RS = transforms.Resize(32)
RS2 = transforms.Resize(output_size)
RC = transforms.RandomCrop(output_size, padding=output_size // 8)
RHF = transforms.RandomHorizontalFlip()
NRM = transforms.Normalize((0.4914, 0.4822, 0.4465),
(0.2023, 0.1994, 0.2010))
TT = transforms.ToTensor()
# Transforms object for trainset with augmentation
transform_with_aug = transforms.Compose([RS, RS2, RC, RHF, TT, NRM])
# Transforms object for testset with NO augmentation
transform_no_aug = transforms.Compose([RS, RS2, TT, NRM])
train_dataset = torchvision.datasets.ImageFolder(
root='./tmp/Food_101_Dataset/food-101/train',
transform=transform_with_aug)
test_dataset = torchvision.datasets.ImageFolder(
root='./tmp/Food_101_Dataset/food-101/test',
transform=transform_no_aug)
train_loader = DataLoader(train_dataset,
batch_size=batch_size,
num_workers=4,
pin_memory=True,
shuffle=True)
test_loader = DataLoader(test_dataset,
batch_size=batch_size,
pin_memory=True,
shuffle=shuffle_test,
num_workers=4)
train_loader.name = "TinyResizedFood101"
return DataBundle(dataset_name="TinyResizedFood101",
train_dataset=train_loader,
test_dataset=test_loader,
cardinality=101,
output_resolution=output_size,
is_classifier=True)
def TinyImageNet(batch_size=12, output_size=64) -> DataBundle:
# Transformations
RC = transforms.RandomCrop(64, padding=64 // 8)
RHF = transforms.RandomHorizontalFlip()
NRM = transforms.Normalize((0.4914, 0.4822, 0.4465),
(0.2023, 0.1994, 0.2010))
TT = transforms.ToTensor()
RS = transforms.Resize(output_size)
# Transforms object for trainset with augmentation
transform_with_aug = transforms.Compose([RC, RHF, RS, TT, NRM])
# Transforms object for testset with NO augmentation
transform_no_aug = transforms.Compose([RS, TT, NRM])
trainset = torchvision.datasets.ImageFolder(
root='./tmp/tiny-imagenet-200/train/', transform=transform_with_aug)
testset = torchvision.datasets.ImageFolder(
root='./tmp/tiny-imagenet-200/val/ds', transform=transform_no_aug)
train_loader = torch.utils.data.DataLoader(trainset,
batch_size=batch_size,
shuffle=True,
num_workers=3,
pin_memory=False)
test_loader = torch.utils.data.DataLoader(testset,
batch_size=batch_size,
shuffle=True,
num_workers=3,
pin_memory=False)
train_loader.name = "TinyImageNet"
return DataBundle(train_dataset=train_loader,
test_dataset=test_loader,
cardinality=200,
output_resolution=output_size,
is_classifier=True,
dataset_name="TinyImageNet")