def test_gradcheck(self, device):
# test parameters
new_size = 4
input = torch.rand(1, 2, 3, 4).to(device)
input = utils.tensor_to_gradcheck_var(input) # to var
assert gradcheck(kornia.Resize(new_size), (input, ),
raise_exception=True)
def get_transform_lib_kornia_resize_crop(in_size=64):
print(get_transform_lib_kornia_resize_crop.__name__)
out_size = 32
return nn.Sequential(
kornia.Resize((48, 48)),
kornia.augmentation.CenterCrop((out_size, out_size)),
).cuda()
def get_transform_AMOS_kornia4(in_size=96, out_size=32):
print(get_transform_AMOS_kornia4.__name__)
return nn.Sequential( # GPU
# kornia.filters.GaussianBlur2d((11, 11), (orig_size / (2. * out_size), orig_size / (2. * out_size))),
torch.nn.ReplicationPad2d(in_size // 4), # otherwise small black corners appear
kornia.augmentation.RandomAffine(degrees=(-25.0, 25.0), scale=(0.7, 1.1), shear=(-20, 20), translate=(0.0, 0.1)),
kornia.augmentation.CenterCrop(in_size),
kornia.Resize((out_size, out_size)),
).cuda()
def get_transform_lib_kornia2ss(in_size=64): # milder
print(get_transform_lib_kornia2ss.__name__)
out_size = 32
return nn.Sequential(
kornia.filters.GaussianBlur2d((7, 7), (0.6, 0.6)), # Blur for proper downscale
torch.nn.ReplicationPad2d(8), # otherwise small black corners appear
kornia.augmentation.RandomAffine(degrees=(0.0, 0.0), scale=(1.0, 1.0), shear=(5.0, 5.0), translate=(0.00, 0.00)),
kornia.augmentation.CenterCrop(in_size),
kornia.Resize((out_size, out_size))
).cuda()
def get_transform_lib_kornia(in_size=64):
print(get_transform_lib_kornia.__name__)
out_size = 32
return nn.Sequential( # GPU
kornia.filters.GaussianBlur2d((11, 11), (in_size / (2. * out_size), in_size / (2. * out_size))),
torch.nn.ReplicationPad2d(in_size // 4), # otherwise small black corners appear
kornia.augmentation.RandomAffine(degrees=(-15.0, 15.0), scale=(0.8, 1.2), shear=(-10, 10), translate=(0.0, 0.07)),
kornia.augmentation.CenterCrop(in_size),
kornia.Resize((out_size, out_size)),
).cuda()
def get_transform_lib_kornia_resize32(in_size=64):
print(get_transform_lib_kornia_resize32.__name__)
out_size = 32
return nn.Sequential(
kornia.Resize((out_size, out_size)),
).cuda()
def get_transform_lib_64(in_size=64): # best for liberty
print(get_transform_lib_64.__name__)
out_size = 64
return nn.Sequential(
kornia.Resize((out_size, out_size))
).cuda()
def test_gradcheck(self, device, dtype):
# test parameters
new_size = 4
input = torch.rand(1, 2, 3, 4, device=device, dtype=dtype)
input = utils.tensor_to_gradcheck_var(input) # to var
assert gradcheck(kornia.Resize(new_size, align_corners=False), (input,), raise_exception=True)
