def test_gaussian(self):
np.testing.assert_allclose(
gaussian_1d(0.5, 8),
torch.tensor(
[
0.0000e00,
2.9802e-07,
1.3496e-03,
1.5731e-01,
6.8269e-01,
1.5731e-01,
1.3496e-03,
2.9802e-07,
0.0000e00,
]
),
rtol=1e-4,
)
np.testing.assert_allclose(
gaussian_1d(1, 1),
torch.tensor([0.24173, 0.382925, 0.24173]),
rtol=1e-4,
)
np.testing.assert_allclose(
gaussian_1d(1, 1, normalize=True),
torch.tensor([0.2790, 0.4420, 0.2790]),
rtol=1e-4,
)
def test_norm_false(self, variance, expected):
extent = 6
atol = 1e-4
sigma = np.sqrt(variance)
k_erf = gaussian_1d(sigma, truncated=extent / sigma, approx="erf", normalize=False).numpy()
k_sampled = gaussian_1d(sigma, truncated=extent / sigma, approx="sampled").numpy()
k_scalespace = gaussian_1d(sigma, truncated=extent / sigma, approx="scalespace").numpy()
np.testing.assert_allclose(k_erf, expected[0], atol=atol)
np.testing.assert_allclose(k_sampled, expected[1], atol=atol)
np.testing.assert_allclose(k_scalespace, expected[2], atol=atol)
def test_scalespace_gaussian(self):
np.testing.assert_allclose(
gaussian_1d(0.5, 8, "scalespace"),
torch.tensor(
[
7.9472e-06,
2.5451e-04,
6.1161e-03,
9.8113e-02,
7.9102e-01,
9.8113e-02,
6.1161e-03,
2.5451e-04,
7.9472e-06,
]
),
rtol=1e-4,
)
np.testing.assert_allclose(
gaussian_1d(1, 1, "scalespace"),
torch.tensor([0.20791, 0.46576, 0.20791]),
rtol=1e-3,
)
np.testing.assert_allclose(
gaussian_1d(1, 1, "scalespace", normalize=True),
torch.tensor([0.2358, 0.5283, 0.2358]),
rtol=1e-3,
)
np.testing.assert_allclose(
gaussian_1d(5, 1, "scalespace"),
torch.tensor(
[
0.048225,
0.057891,
0.06675,
0.073911,
0.078576,
0.080197,
0.078576,
0.073911,
0.06675,
0.057891,
0.048225,
]
),
rtol=1e-3,
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Args:
x: in shape [Batch, chns, H, W, D].
"""
_kernel = [gaussian_1d(s, truncated=self.truncated, approx=self.approx) for s in self.sigma]
return separable_filtering(x=x, kernels=_kernel)
def __init__(self,
spatial_dims: int,
sigma: Union[Sequence[float], float],
truncated: float = 4.0) -> None:
"""
Args:
spatial_dims: number of spatial dimensions of the input image.
must have shape (Batch, channels, H[, W, ...]).
sigma: std.
truncated: spreads how many stds.
"""
super().__init__()
self.spatial_dims = int(spatial_dims)
_sigma = ensure_tuple_rep(sigma, self.spatial_dims)
self.kernel = [
torch.nn.Parameter(
torch.as_tensor(gaussian_1d(s, truncated), dtype=torch.float),
False) for s in _sigma
]
self.padding = [
cast(int, (same_padding(k.size()[0]))) for k in self.kernel
]
self.conv_n = [F.conv1d, F.conv2d, F.conv3d][spatial_dims - 1]
for idx, param in enumerate(self.kernel):
self.register_parameter(f"kernel_{idx}", param)
def __init__(self, spatial_dims, sigma, truncated=4., device=None):
"""
Args:
spatial_dims (int): number of spatial dimensions of the input image.
must have shape (Batch, channels, H[, W, ...]).
sigma (float): std.
truncated (float): spreads how many stds.
device (torch.device): device on which the tensor will be allocated.
"""
self.kernel = torch.nn.Parameter(torch.tensor(gaussian_1d(sigma, truncated)), False)
self.spatial_dims = spatial_dims
self.conv_n = [F.conv1d, F.conv2d, F.conv3d][spatial_dims - 1]
self.padding = same_padding(self.kernel.size()[0])
self.device = device
self.kernel = self.kernel.to(self.device)
def test_wrong_sigma(self):
with self.assertRaises(ValueError):
gaussian_1d(-1, 10)
with self.assertRaises(ValueError):
gaussian_1d(1, -10)
def test_wrong_sigma(self):
with self.assertRaises(ValueError):
gaussian_1d(1, -10)
with self.assertRaises(NotImplementedError):
gaussian_1d(1, 10, "wrong_arg")
