def sample_and_log_prob(self, num_samples, context=None):
"""Generates samples from the flow, together with their log probabilities.
For flows, this is more efficient that calling `sample` and `log_prob` separately.
"""
embedded_context = self._embedding_net(context)
noise, log_prob = self._distribution.sample_and_log_prob(
num_samples, context=embedded_context)
if embedded_context is not None:
# Merge the context dimension with sample dimension in order to apply the transform.
noise = torchutils.merge_leading_dims(noise, num_dims=2)
embedded_context = torchutils.repeat_rows(embedded_context,
num_reps=num_samples)
samples, logabsdet = self._transform.inverse(noise,
context=embedded_context)
if embedded_context is not None:
# Split the context dimension from sample dimension.
samples = torchutils.split_leading_dim(samples,
shape=[-1, num_samples])
logabsdet = torchutils.split_leading_dim(logabsdet,
shape=[-1, num_samples])
return samples, log_prob - logabsdet
def sample_and_log_prob(self, num_samples, context=None):
"""Generates samples from the distribution together with their log probability.
Args:
num_samples: int, number of samples to generate.
context: Tensor or None, conditioning variables. If None, the context is ignored.
Returns:
A tuple of:
* A Tensor containing the samples, with shape [num_samples, ...] if context is None,
or [context_size, num_samples, ...] if context is given.
* A Tensor containing the log probabilities of the samples, with shape
[num_samples, ...] if context is None, or [context_size, num_samples, ...] if
context is given.
"""
samples = self.sample(num_samples, context=context)
if context is not None:
# Merge the context dimension with sample dimension in order to call log_prob.
samples = torchutils.merge_leading_dims(samples, num_dims=2)
context = torchutils.repeat_rows(context, num_reps=num_samples)
assert samples.shape[0] == context.shape[0]
log_prob = self.log_prob(samples, context=context)
if context is not None:
# Split the context dimension from sample dimension.
samples = torchutils.split_leading_dim(samples,
shape=[-1, num_samples])
log_prob = torchutils.split_leading_dim(log_prob,
shape=[-1, num_samples])
return samples, log_prob
def _sample(self, num_samples, context):
if context is None:
return torch.randn(num_samples, *self._shape)
else:
# The value of the context is ignored, only its size is taken into account.
context_size = context.shape[0]
samples = torch.randn(context_size * num_samples, *self._shape)
return torchutils.split_leading_dim(samples,
[context_size, num_samples])
def _sample(self, num_samples, context):
# Compute parameters.
logits = self._compute_params(context)
probs = torch.sigmoid(logits)
probs = torchutils.repeat_rows(probs, num_samples)
# Generate samples.
context_size = context.shape[0]
noise = torch.rand(context_size * num_samples, *self._shape)
samples = (noise < probs).float()
return torchutils.split_leading_dim(samples,
[context_size, num_samples])
def _sample(self, num_samples, context):
# Compute parameters.
means, log_stds = self._compute_params(context)
stds = torch.exp(log_stds)
means = torchutils.repeat_rows(means, num_samples)
stds = torchutils.repeat_rows(stds, num_samples)
# Generate samples.
context_size = context.shape[0]
noise = torch.randn(context_size * num_samples, *self._shape)
samples = means + stds * noise
return torchutils.split_leading_dim(samples,
[context_size, num_samples])
def _sample(self, num_samples, context):
embedded_context = self._embedding_net(context)
noise = self._distribution.sample(num_samples,
context=embedded_context)
if embedded_context is not None:
# Merge the context dimension with sample dimension in order to apply the transform.
noise = torchutils.merge_leading_dims(noise, num_dims=2)
embedded_context = torchutils.repeat_rows(embedded_context,
num_reps=num_samples)
samples, _ = self._transform.inverse(noise, context=embedded_context)
if embedded_context is not None:
# Split the context dimension from sample dimension.
samples = torchutils.split_leading_dim(samples,
shape=[-1, num_samples])
return samples