def create_transform(self, shape, split_dim=1):
mct = base.MultiscaleCompositeTransform(num_transforms=4,
split_dim=split_dim)
for transform in [
standard.AffineScalarTransform(scale=2.),
standard.AffineScalarTransform(scale=4.),
standard.AffineScalarTransform(scale=0.5),
standard.AffineScalarTransform(scale=0.25)
]:
shape = mct.add_transform(transform, shape)
return mct
def test_inverse(self):
batch_size = 10
shape = [2, 3, 4]
inputs = torch.randn(batch_size, *shape)
transform = base.InverseTransform(
standard.AffineScalarTransform(scale=-2.0))
reference = standard.AffineScalarTransform(scale=-0.5)
outputs, logabsdet = transform.inverse(inputs)
outputs_ref, logabsdet_ref = reference.inverse(inputs)
self.assert_tensor_is_good(outputs, [batch_size] + shape)
self.assert_tensor_is_good(logabsdet, [batch_size])
self.assertEqual(outputs, outputs_ref)
self.assertEqual(logabsdet, logabsdet_ref)
def test_inverse(self):
batch_size = 10
shape = [2, 3, 4]
inputs = torch.randn(batch_size, *shape)
transforms = [
standard.AffineScalarTransform(scale=2.0),
standard.IdentityTransform(),
standard.AffineScalarTransform(scale=-0.25),
]
composite = base.CompositeTransform(transforms)
reference = standard.AffineScalarTransform(scale=-0.5)
outputs, logabsdet = composite.inverse(inputs)
outputs_ref, logabsdet_ref = reference.inverse(inputs)
self.assert_tensor_is_good(outputs, [batch_size] + shape)
self.assert_tensor_is_good(logabsdet, [batch_size])
self.assertEqual(outputs, outputs_ref)
self.assertEqual(logabsdet, logabsdet_ref)
def test_case(scale, shift, true_outputs, true_logabsdet):
with self.subTest(scale=scale, shift=shift):
transform = standard.AffineScalarTransform(scale=scale, shift=shift)
outputs, logabsdet = transform.inverse(inputs)
self.assert_tensor_is_good(outputs, [batch_size] + shape)
self.assert_tensor_is_good(logabsdet, [batch_size])
self.assertEqual(outputs, true_outputs)
self.assertEqual(logabsdet,
torch.full([batch_size], true_logabsdet * np.prod(shape)))
def test_case(scale, shift):
transform = standard.AffineScalarTransform(scale=scale, shift=shift)
self.assert_forward_inverse_are_consistent(transform, inputs)