def standardizing_transform(
batch_t: Tensor, min_std: float = 1e-14
) -> transforms.AffineTransform:
"""Builds standardizing transform
Args:
batch_t: Batched tensor from which mean and std deviation (across
first dimension) are computed.
min_std: Minimum value of the standard deviation to use when z-scoring to
avoid division by zero.
Returns:
Affine transform for z-scoring
"""
is_valid_t, *_ = handle_invalid_x(batch_t, True)
t_mean = torch.mean(batch_t[is_valid_t], dim=0)
t_std = torch.std(batch_t[is_valid_t], dim=0)
t_std[t_std < min_std] = min_std
return transforms.AffineTransform(shift=-t_mean / t_std, scale=1 / t_std)
def standardizing_transform(
batch_t: Tensor,
structured_dims: bool = False,
min_std: float = 1e-14) -> transforms.AffineTransform:
"""Builds standardizing transform
Args:
batch_t: Batched tensor from which mean and std deviation (across
first dimension) are computed.
structured_dim: Whether data dimensions are structured (e.g., time-series,
images), which requires computing mean and std per sample first before
aggregating over samples for a single standardization mean and std for the
batch, or independent (default), which z-scores dimensions independently.
min_std: Minimum value of the standard deviation to use when z-scoring to
avoid division by zero.
Returns:
Affine transform for z-scoring
"""
is_valid_t, *_ = handle_invalid_x(batch_t, True)
if structured_dims:
# Structured data so compute a single mean over all dimensions
# equivalent to taking mean over per-sample mean, i.e.,
# `torch.mean(torch.mean(.., dim=1))`.
t_mean = torch.mean(batch_t[is_valid_t])
# Compute std per-sample first.
sample_std = torch.std(batch_t[is_valid_t], dim=1)
sample_std[sample_std < min_std] = min_std
# Average over all samples for batch std.
t_std = torch.mean(sample_std)
else:
t_mean = torch.mean(batch_t[is_valid_t], dim=0)
t_std = torch.std(batch_t[is_valid_t], dim=0)
t_std[t_std < min_std] = min_std
return transforms.AffineTransform(shift=-t_mean / t_std, scale=1 / t_std)
def posterior_nn(
model: str,
prior: torch.distributions.Distribution,
context: torch.Tensor,
embedding: Optional[torch.nn.Module] = None,
hidden_features: int = 50,
mdn_num_components: int = 20,
made_num_mixture_components: int = 10,
made_num_blocks: int = 4,
flow_num_transforms: int = 5,
) -> torch.nn.Module:
"""Neural posterior density estimator
Args:
model: Model, one of maf / mdn / made / nsf
prior: Prior distribution
context: Observation
embedding: Embedding network
hidden_features: For all, number of hidden features
mdn_num_components: For MDNs only, number of components
made_num_mixture_components: For MADEs only, number of mixture components
made_num_blocks: For MADEs only, number of blocks
flow_num_transforms: For flows only, number of transforms
Returns:
Neural network
"""
mean, std = (prior.mean, prior.stddev)
standardizing_transform = transforms.AffineTransform(
shift=-mean / std, scale=1 / std
)
parameter_dim = prior.sample([1]).shape[1]
context = utils.torchutils.atleast_2d(context)
observation_dim = torch.tensor([context.shape[1:]])
if model == "mdn":
neural_net = MultivariateGaussianMDN(
features=parameter_dim,
context_features=observation_dim,
hidden_features=hidden_features,
hidden_net=nn.Sequential(
nn.Linear(observation_dim, hidden_features),
nn.ReLU(),
nn.Dropout(p=0.0),
nn.Linear(hidden_features, hidden_features),
nn.ReLU(),
nn.Linear(hidden_features, hidden_features),
nn.ReLU(),
),
num_components=mdn_num_components,
custom_initialization=True,
)
elif model == "made":
transform = standardizing_transform
distribution = distributions_.MADEMoG(
features=parameter_dim,
hidden_features=hidden_features,
context_features=observation_dim,
num_blocks=made_num_blocks,
num_mixture_components=made_num_mixture_components,
use_residual_blocks=True,
random_mask=False,
activation=torch.relu,
dropout_probability=0.0,
use_batch_norm=False,
custom_initialization=True,
)
neural_net = flows.Flow(transform, distribution, embedding)
elif model == "maf":
transform = transforms.CompositeTransform(
[
transforms.CompositeTransform(
[
transforms.MaskedAffineAutoregressiveTransform(
features=parameter_dim,
hidden_features=hidden_features,
context_features=observation_dim,
num_blocks=2,
use_residual_blocks=False,
random_mask=False,
activation=torch.tanh,
dropout_probability=0.0,
use_batch_norm=True,
),
transforms.RandomPermutation(features=parameter_dim),
]
)
for _ in range(flow_num_transforms)
]
)
transform = transforms.CompositeTransform([standardizing_transform, transform,])
distribution = distributions_.StandardNormal((parameter_dim,))
neural_net = flows.Flow(transform, distribution, embedding)
elif model == "nsf":
transform = transforms.CompositeTransform(
[
transforms.CompositeTransform(
[
transforms.PiecewiseRationalQuadraticCouplingTransform(
mask=create_alternating_binary_mask(
features=parameter_dim, even=(i % 2 == 0)
),
transform_net_create_fn=lambda in_features, out_features: nets.ResidualNet(
in_features=in_features,
out_features=out_features,
hidden_features=hidden_features,
context_features=observation_dim,
num_blocks=2,
activation=torch.relu,
dropout_probability=0.0,
use_batch_norm=False,
),
num_bins=10,
tails="linear",
tail_bound=3.0,
apply_unconditional_transform=False,
),
transforms.LULinear(parameter_dim, identity_init=True),
]
)
for i in range(flow_num_transforms)
]
)
transform = transforms.CompositeTransform([standardizing_transform, transform,])
distribution = distributions_.StandardNormal((parameter_dim,))
neural_net = flows.Flow(transform, distribution, embedding)
else:
raise ValueError
return neural_net
