def test_exception(self, device, dtype):
with pytest.raises(TypeError):
assert elastic_transform2d([0.])
with pytest.raises(TypeError):
assert elastic_transform2d(torch.tensor(), 1)
with pytest.raises(ValueError):
img = torch.ones(1, 1, 1, device=device, dtype=dtype)
noise = torch.ones(1, 2, 1, 1, device=device, dtype=dtype)
assert elastic_transform2d(img, noise)
with pytest.raises(ValueError):
img = torch.ones(1, 1, 1, 1, device=device, dtype=dtype)
noise = torch.ones(1, 3, 1, 1, device=device, dtype=dtype)
assert elastic_transform2d(img, noise)
def test_cardinality(self, device, dtype, batch, channels, height, width):
shape = batch, channels, height, width
img = torch.ones(shape, device=device, dtype=dtype)
noise = torch.ones((batch, 2, height, width),
device=device,
dtype=dtype)
assert elastic_transform2d(img, noise).shape == shape
def inner(x):
a = np.random.rand(2)
k = np.random.randint(8,64) * 2 + 1 # 63
s = k / (np.random.rand()+2.) # 2-3 times less than k
# s = float(np.random.randint(8,64)) # 32
noise = torch.zeros([1, 2, x.shape[2], x.shape[3]]).cuda()
return K.elastic_transform2d(x, noise, (k,k), (s,s), tuple(a))
def test_valid_paramters(self, device, dtype, kernel_size, sigma, alpha):
image = torch.rand(1, 4, 5, 5, device=device, dtype=dtype)
noise = torch.rand(1, 2, 5, 5, device=device, dtype=dtype)
if isinstance(sigma, torch.Tensor):
sigma = sigma.to(device, dtype)
if isinstance(alpha, torch.Tensor):
alpha = alpha.to(device, dtype)
assert elastic_transform2d(image, noise, kernel_size, sigma,
alpha) is not None
def apply_transform(self,
input: Tensor,
params: Dict[str, Tensor],
transform: Optional[Tensor] = None) -> Tensor:
return elastic_transform2d(
input,
params["noise"].to(input),
self.flags["kernel_size"],
self.flags["sigma"],
self.flags["alpha"],
self.flags["align_corners"],
self.flags["mode"],
self.flags["padding_mode"],
)
def test_values(self, device, dtype):
image = torch.tensor(
[[[[0.0018, 0.7521, 0.7550], [0.2053, 0.4249, 0.1369],
[0.1027, 0.3992, 0.8773]]]],
device=device,
dtype=dtype)
noise = torch.ones(1, 2, 3, 3, device=device, dtype=dtype)
expected = torch.tensor(
[[[[0.2193, 0.2193, 0.2193], [0.2193, 0.2193, 0.2193],
[0.2193, 0.2193, 0.2193]]]],
device=device,
dtype=dtype)
actual = elastic_transform2d(image, noise)
assert_allclose(actual, expected, atol=1e-3, rtol=1e-3)
def test_smoke(self, device, dtype):
image = torch.rand(1, 4, 5, 5, device=device, dtype=dtype)
noise = torch.rand(1, 2, 5, 5, device=device, dtype=dtype)
assert elastic_transform2d(image, noise) is not None