def __init__(self,
data_dim,
latent_dim,
encoder,
outer_transform,
inner_transform=None,
pie_epsilon=1.0e-2,
apply_context_to_outer=True):
super(EncoderManifoldFlow, self).__init__()
assert latent_dim < data_dim
self.data_dim = data_dim
self.latent_dim = latent_dim
self.total_data_dim = product(data_dim)
self.total_latent_dim = product(latent_dim)
self.apply_context_to_outer = apply_context_to_outer
self.manifold_latent_distribution = distributions.StandardNormal(
(self.total_latent_dim, ))
self.orthogonal_latent_distribution = distributions.RescaledNormal(
(self.total_data_dim - self.total_latent_dim, ),
std=pie_epsilon,
clip=5.0 * pie_epsilon)
self.projection = ProjectionSplit(self.total_data_dim,
self.total_latent_dim)
self.encoder = encoder
self.outer_transform = outer_transform
if inner_transform is None:
self.inner_transform = transforms.IdentityTransform()
else:
self.inner_transform = inner_transform
self._report_model_parameters()
def __init__(self, data_dim, transform):
super(Flow, self).__init__()
self.data_dim = data_dim
self.latent_dim = data_dim
self.total_data_dim = product(data_dim)
self.total_latent_dim = product(self.latent_dim)
self.latent_distribution = distributions.StandardNormal(
(self.total_latent_dim, ))
self.transform = transform
self._report_model_parameters()
def __init__(self, data_dim, latent_dim, encoder, decoder, inner_transform=None, apply_context_to_outer=True):
super(ProbabilisticAutoEncoder, self).__init__()
self.data_dim = data_dim
self.latent_dim = latent_dim
self.total_data_dim = product(data_dim)
self.total_latent_dim = product(latent_dim)
self.apply_context_to_outer = apply_context_to_outer
assert self.total_latent_dim < self.total_data_dim
self.manifold_latent_distribution = distributions.StandardNormal((self.total_latent_dim,))
self.encoder = encoder
self.decoder = decoder
if inner_transform is None:
self.inner_transform = transforms.IdentityTransform()
else:
self.inner_transform = inner_transform
self._report_model_parameters()