def logabsdet(self):
"""Cost:
logabsdet = O(D^3)
where:
D = num of features
"""
return torchutils.logabsdet(self._weight)
def test_logabsdet(self):
size = 10
matrix = torch.randn(size, size)
logabsdet = torchutils.logabsdet(matrix)
logabsdet_ref = torch.log(torch.abs(matrix.det()))
self.eps = 1e-6
self.assertEqual(logabsdet, logabsdet_ref)
def setUp(self):
self.features = 3
self.transform = linear.NaiveLinear(features=self.features)
self.weight = self.transform._weight
self.weight_inverse = torch.inverse(self.weight)
self.logabsdet = torchutils.logabsdet(self.weight)
self.eps = 1e-5
def setUp(self):
self.features = 3
self.transform = lu.LULinear(features=self.features)
lower, upper = self.transform._create_lower_upper()
self.weight = lower @ upper
self.weight_inverse = torch.inverse(self.weight)
self.logabsdet = torchutils.logabsdet(self.weight)
self.eps = 1e-5
def setUp(self):
self.features = 3
self.transform = qr.QRLinear(features=self.features, num_householder=4)
upper = self.transform._create_upper()
orthogonal = self.transform.orthogonal.matrix()
self.weight = orthogonal @ upper
self.weight_inverse = torch.inverse(self.weight)
self.logabsdet = torchutils.logabsdet(self.weight)
self.eps = 1e-5
def forward_no_cache(self, inputs):
"""Cost:
output = O(D^2N)
logabsdet = O(D^3)
where:
D = num of features
N = num of inputs
"""
batch_size = inputs.shape[0]
outputs = F.linear(inputs, self._weight, self.bias)
logabsdet = torchutils.logabsdet(self._weight)
logabsdet = logabsdet * torch.ones(batch_size)
return outputs, logabsdet
def setUp(self):
self.features = 3
self.transform = SVDLinear(features=self.features, num_householder=4)
self.transform.bias.data = torch.randn(
self.features) # Just so bias isn't zero.
diagonal = torch.diag(torch.exp(self.transform.log_diagonal))
orthogonal_1 = self.transform.orthogonal_1.matrix()
orthogonal_2 = self.transform.orthogonal_2.matrix()
self.weight = orthogonal_1 @ diagonal @ orthogonal_2
self.weight_inverse = torch.inverse(self.weight)
self.logabsdet = torchutils.logabsdet(self.weight)
self.eps = 1e-5
def test_inverse_full_orthogonal(self):
features = 100
batch_size = 50
transform = orthogonal.FullOrthogonalTransform(features=features)
matrix = transform.matrix()
inputs = torch.randn(batch_size, features)
outputs, logabsdet = transform.inverse(inputs)
self.assert_tensor_is_good(outputs, [batch_size, features])
self.assert_tensor_is_good(logabsdet, [batch_size])
self.eps = 1e-5
self.assertEqual(outputs, inputs @ matrix)
self.assertEqual(logabsdet,
torchutils.logabsdet(matrix) * torch.ones(batch_size))
self.assert_forward_inverse_are_consistent(transform, inputs)
def test_inverse(self):
features = 100
batch_size = 50
for num_transforms in [1, 2, 11, 12]:
with self.subTest(num_transforms=num_transforms):
transform = orthogonal.HouseholderSequence(
features=features, num_transforms=num_transforms)
matrix = transform.matrix()
inputs = torch.randn(batch_size, features)
outputs, logabsdet = transform.inverse(inputs)
self.assert_tensor_is_good(outputs, [batch_size, features])
self.assert_tensor_is_good(logabsdet, [batch_size])
self.eps = 1e-5
self.assertEqual(outputs, inputs @ matrix)
self.assertEqual(
logabsdet,
torchutils.logabsdet(matrix) * torch.ones(batch_size))