def test_forward_inverse_are_consistent(self):
batch_size = 10
features = 20
inputs = torch.randn(batch_size, features)
self.eps = 1e-6
for use_residual_blocks, random_mask in [
(False, False),
(False, True),
(True, False),
]:
with self.subTest(use_residual_blocks=use_residual_blocks,
random_mask=random_mask):
transform = autoregressive.MaskedAffineAutoregressiveTransform(
features=features,
hidden_features=30,
num_blocks=5,
use_residual_blocks=use_residual_blocks,
random_mask=random_mask,
)
self.assert_forward_inverse_are_consistent(transform, inputs)
def test_inverse(self):
batch_size = 10
features = 20
inputs = torch.randn(batch_size, features)
for use_residual_blocks, random_mask in [
(False, False),
(False, True),
(True, False),
]:
with self.subTest(use_residual_blocks=use_residual_blocks,
random_mask=random_mask):
transform = autoregressive.MaskedAffineAutoregressiveTransform(
features=features,
hidden_features=30,
num_blocks=5,
use_residual_blocks=use_residual_blocks,
random_mask=random_mask,
)
outputs, logabsdet = transform.inverse(inputs)
self.assert_tensor_is_good(outputs, [batch_size, features])
self.assert_tensor_is_good(logabsdet, [batch_size])