def test_expected_shapes():
val = torch.zeros((2, 3, 4, 5))
u, s, vh, _ = decompositions.svd_decomposition(torch, val, 2)
assert u.shape == (2, 3, 6)
assert s.shape == (6, )
np.testing.assert_allclose(s, np.zeros(6))
assert vh.shape == (6, 4, 5)
def test_max_truncation_error_relative():
absolute = np.diag([2.0, 1.0, 0.2, 0.1])
relative = np.diag([2.0, 1.0, 0.2, 0.1])
max_truncation_err = 0.2
_, _, _, trunc_sv_absolute = decompositions.svd_decomposition(
torch,
torch.Tensor(absolute),
1,
max_truncation_error=max_truncation_err,
relative=False)
_, _, _, trunc_sv_relative = decompositions.svd_decomposition(
torch,
torch.Tensor(relative),
1,
max_truncation_error=max_truncation_err,
relative=True)
np.testing.assert_almost_equal(trunc_sv_absolute, [0.1])
np.testing.assert_almost_equal(trunc_sv_relative, [0.2, 0.1])
def svd_decomposition(self,
tensor: Tensor,
split_axis: int,
max_singular_values: Optional[int] = None,
max_truncation_error: Optional[float] = None
) -> Tuple[Tensor, Tensor, Tensor, Tensor]:
return decompositions.svd_decomposition(self.torch, tensor, split_axis,
max_singular_values,
max_truncation_error)
def test_max_truncation_error():
np.random.seed(2019)
random_matrix = np.random.rand(10, 10)
unitary1, _, unitary2 = np.linalg.svd(random_matrix)
singular_values = np.array(range(10))
val = unitary1.dot(np.diag(singular_values).dot(unitary2.T))
u, s, vh, trun = decompositions.svd_decomposition(
torch, torch.Tensor(val), 1, max_truncation_error=math.sqrt(5.1))
assert u.shape == (10, 7)
assert s.shape == (7, )
np.testing.assert_array_almost_equal(s, np.arange(9, 2, -1), decimal=5)
assert vh.shape == (7, 10)
np.testing.assert_array_almost_equal(trun, np.arange(2, -1, -1))
def svd_decomposition(
self,
tensor: Tensor,
split_axis: int,
max_singular_values: Optional[int] = None,
max_truncation_error: Optional[float] = None,
relative: Optional[bool] = False
) -> Tuple[Tensor, Tensor, Tensor, Tensor]:
return decompositions.svd_decomposition(torchlib,
tensor,
split_axis,
max_singular_values,
max_truncation_error,
relative=relative)
