def inverse(self, inputs, context=None):
if torch.min(inputs) <= -1 or torch.max(inputs) >= 1:
raise InputOutsideDomain()
outputs = 0.5 * torch.log((1 + inputs) / (1 - inputs))
logabsdet = -torch.log(1 - inputs ** 2)
logabsdet = torchutils.sum_except_batch(logabsdet, num_batch_dims=1)
return outputs, logabsdet
def forward(self, inputs, context=None):
inputs = self.temperature * inputs
outputs = torch.sigmoid(inputs)
logabsdet = torchutils.sum_except_batch(
torch.log(self.temperature) - F.softplus(-inputs) - F.softplus(inputs)
)
return outputs, logabsdet
def _spline(self, inputs, inverse=False):
batch_size = inputs.shape[0]
unnormalized_widths = _share_across_batch(self.unnormalized_widths, batch_size)
unnormalized_heights = _share_across_batch(
self.unnormalized_heights, batch_size
)
unnormalized_derivatives = _share_across_batch(
self.unnormalized_derivatives, batch_size
)
if self.tails is None:
spline_fn = splines.rational_quadratic_spline
spline_kwargs = {}
else:
spline_fn = splines.unconstrained_rational_quadratic_spline
spline_kwargs = {"tails": self.tails, "tail_bound": self.tail_bound}
outputs, logabsdet = spline_fn(
inputs=inputs,
unnormalized_widths=unnormalized_widths,
unnormalized_heights=unnormalized_heights,
unnormalized_derivatives=unnormalized_derivatives,
inverse=inverse,
min_bin_width=self.min_bin_width,
min_bin_height=self.min_bin_height,
min_derivative=self.min_derivative,
**spline_kwargs
)
return outputs, torchutils.sum_except_batch(logabsdet)
def _elementwise(self, inputs, autoregressive_params, inverse=False):
batch_size = inputs.shape[0]
transform_params = autoregressive_params.view(batch_size,
self.features,
self.num_bins * 2 + 2)
unnormalized_widths = transform_params[..., :self.num_bins]
unnormalized_heights = transform_params[..., self.num_bins:2 *
self.num_bins]
derivatives = transform_params[..., 2 * self.num_bins:]
unnorm_derivatives_left = derivatives[..., 0][..., None]
unnorm_derivatives_right = derivatives[..., 1][..., None]
if hasattr(self.autoregressive_net, "hidden_features"):
unnormalized_widths /= np.sqrt(
self.autoregressive_net.hidden_features)
unnormalized_heights /= np.sqrt(
self.autoregressive_net.hidden_features)
outputs, logabsdet = cubic_spline(
inputs=inputs,
unnormalized_widths=unnormalized_widths,
unnormalized_heights=unnormalized_heights,
unnorm_derivatives_left=unnorm_derivatives_left,
unnorm_derivatives_right=unnorm_derivatives_right,
inverse=inverse,
)
return outputs, torchutils.sum_except_batch(logabsdet)
def _log_prob(self, inputs, context):
if inputs.shape[1:] != self._shape:
raise ValueError("Expected input of shape {}, got {}".format(
self._shape, inputs.shape[1:]))
# Compute parameters.
means, log_stds = self._compute_params(context)
assert means.shape == inputs.shape and log_stds.shape == inputs.shape
# Compute log prob.
norm_inputs = (inputs - means) * torch.exp(-log_stds)
log_prob = -0.5 * torchutils.sum_except_batch(norm_inputs**2,
num_batch_dims=1)
log_prob -= torchutils.sum_except_batch(log_stds, num_batch_dims=1)
log_prob -= self._log_z
return log_prob
def _log_prob(self, inputs, context):
# Note: the context is ignored.
if inputs.shape[1:] != self._shape:
raise ValueError("Expected input of shape {}, got {}".format(
self._shape, inputs.shape[1:]))
neg_energy = -0.5 * torchutils.sum_except_batch(inputs**2,
num_batch_dims=1)
return neg_energy - self._log_z
def _elementwise_forward(self, inputs, autoregressive_params):
unconstrained_scale, shift = self._unconstrained_scale_and_shift(
autoregressive_params)
# scale = torch.sigmoid(unconstrained_scale + 2.0) + self._epsilon
scale = F.softplus(unconstrained_scale) + self._epsilon
log_scale = torch.log(scale)
outputs = scale * inputs + shift
logabsdet = torchutils.sum_except_batch(log_scale, num_batch_dims=1)
return outputs, logabsdet
def inverse(self, inputs, context=None):
if torch.min(inputs) < 0 or torch.max(inputs) > 1:
raise InputOutsideDomain()
outputs = torch.tan(np.pi * (inputs - 0.5))
logabsdet = -torchutils.sum_except_batch(
-np.log(np.pi) - torch.log(1 + outputs ** 2)
)
return outputs, logabsdet
def _elementwise(self, inputs, autoregressive_params, inverse=False):
batch_size = inputs.shape[0]
unnormalized_pdf = autoregressive_params.view(
batch_size, self.features, self._output_dim_multiplier())
outputs, logabsdet = linear_spline(inputs=inputs,
unnormalized_pdf=unnormalized_pdf,
inverse=inverse)
return outputs, torchutils.sum_except_batch(logabsdet)
def inverse(self, inputs, context=None):
if torch.min(inputs) < 0 or torch.max(inputs) > 1:
raise InputOutsideDomain()
inputs = torch.clamp(inputs, self.eps, 1 - self.eps)
outputs = (1 / self.temperature) * (torch.log(inputs) - torch.log1p(-inputs))
logabsdet = -torchutils.sum_except_batch(
torch.log(self.temperature)
- F.softplus(-self.temperature * outputs)
- F.softplus(self.temperature * outputs)
)
return outputs, logabsdet
def _lu_forward_inverse(self, inputs, inverse=False):
b, c, h, w = inputs.shape
inputs = inputs.permute(0, 2, 3, 1).reshape(b * h * w, c)
if inverse:
outputs, logabsdet = super().inverse(inputs)
else:
outputs, logabsdet = super().forward(inputs)
outputs = outputs.reshape(b, h, w, c).permute(0, 3, 1, 2)
logabsdet = logabsdet.reshape(b, h, w)
return outputs, torchutils.sum_except_batch(logabsdet)
def _log_prob(self, inputs, context):
if inputs.shape[1:] != self._shape:
raise ValueError("Expected input of shape {}, got {}".format(
self._shape, inputs.shape[1:]))
# Compute parameters.
logits = self._compute_params(context)
assert logits.shape == inputs.shape
# Compute log prob.
log_prob = -inputs * F.softplus(-logits) - (
1.0 - inputs) * F.softplus(logits)
log_prob = torchutils.sum_except_batch(log_prob, num_batch_dims=1)
return log_prob
def forward(self, inputs, context=None):
mask_right = inputs > self.cut_point
mask_left = inputs < -self.cut_point
mask_middle = ~(mask_right | mask_left)
outputs = torch.zeros_like(inputs)
outputs[mask_middle] = torch.tanh(inputs[mask_middle])
outputs[mask_right] = self.alpha * torch.log(self.beta * inputs[mask_right])
outputs[mask_left] = self.alpha * -torch.log(-self.beta * inputs[mask_left])
logabsdet = torch.zeros_like(inputs)
logabsdet[mask_middle] = torch.log(1 - outputs[mask_middle] ** 2)
logabsdet[mask_right] = torch.log(self.alpha / inputs[mask_right])
logabsdet[mask_left] = torch.log(-self.alpha / inputs[mask_left])
logabsdet = torchutils.sum_except_batch(logabsdet, num_batch_dims=1)
return outputs, logabsdet
def _spline(self, inputs, inverse=False):
batch_size = inputs.shape[0]
unnormalized_pdf = _share_across_batch(self.unnormalized_pdf, batch_size)
if self.tails is None:
outputs, logabsdet = splines.linear_spline(
inputs=inputs, unnormalized_pdf=unnormalized_pdf, inverse=inverse
)
else:
outputs, logabsdet = splines.unconstrained_linear_spline(
inputs=inputs,
unnormalized_pdf=unnormalized_pdf,
inverse=inverse,
tails=self.tails,
tail_bound=self.tail_bound,
)
return outputs, torchutils.sum_except_batch(logabsdet)
def _elementwise(self, inputs, autoregressive_params, inverse=False):
batch_size, features = inputs.shape[0], inputs.shape[1]
transform_params = autoregressive_params.view(
batch_size, features, self._output_dim_multiplier())
unnormalized_widths = transform_params[..., :self.num_bins]
unnormalized_heights = transform_params[..., self.num_bins:2 *
self.num_bins]
unnormalized_derivatives = transform_params[..., 2 * self.num_bins:]
if hasattr(self.autoregressive_net, "hidden_features"):
unnormalized_widths /= np.sqrt(
self.autoregressive_net.hidden_features)
unnormalized_heights /= np.sqrt(
self.autoregressive_net.hidden_features)
if self.tails is None:
spline_fn = rational_quadratic_spline
spline_kwargs = {}
elif self.tails == "linear":
spline_fn = unconstrained_rational_quadratic_spline
spline_kwargs = {
"tails": self.tails,
"tail_bound": self.tail_bound
}
else:
raise ValueError
outputs, logabsdet = spline_fn(
inputs=inputs,
unnormalized_widths=unnormalized_widths,
unnormalized_heights=unnormalized_heights,
unnormalized_derivatives=unnormalized_derivatives,
inverse=inverse,
min_bin_width=self.min_bin_width,
min_bin_height=self.min_bin_height,
min_derivative=self.min_derivative,
**spline_kwargs)
return outputs, torchutils.sum_except_batch(logabsdet)
def inverse(self, inputs, context=None):
mask_right = inputs > self.inv_cut_point
mask_left = inputs < -self.inv_cut_point
mask_middle = ~(mask_right | mask_left)
outputs = torch.zeros_like(inputs)
outputs[mask_middle] = 0.5 * torch.log(
(1 + inputs[mask_middle]) / (1 - inputs[mask_middle])
)
outputs[mask_right] = torch.exp(inputs[mask_right] / self.alpha) / self.beta
outputs[mask_left] = -torch.exp(-inputs[mask_left] / self.alpha) / self.beta
logabsdet = torch.zeros_like(inputs)
logabsdet[mask_middle] = -torch.log(1 - inputs[mask_middle] ** 2)
logabsdet[mask_right] = (
-np.log(self.alpha * self.beta) + inputs[mask_right] / self.alpha
)
logabsdet[mask_left] = (
-np.log(self.alpha * self.beta) - inputs[mask_left] / self.alpha
)
logabsdet = torchutils.sum_except_batch(logabsdet, num_batch_dims=1)
return outputs, logabsdet
def inverse(self, inputs, context=None):
outputs = F.leaky_relu(inputs, negative_slope=(1 / self.negative_slope))
mask = (inputs < 0).type(torch.Tensor)
logabsdet = -self.log_negative_slope * mask
logabsdet = torchutils.sum_except_batch(logabsdet, num_batch_dims=1)
return outputs, logabsdet
def forward(self, inputs, context=None):
outputs = (1 / np.pi) * torch.atan(inputs) + 0.5
logabsdet = torchutils.sum_except_batch(
-np.log(np.pi) - torch.log(1 + inputs ** 2)
)
return outputs, logabsdet
def forward(self, inputs, context=None):
outputs = torch.tanh(inputs)
logabsdet = torch.log(1 - outputs ** 2)
logabsdet = torchutils.sum_except_batch(logabsdet, num_batch_dims=1)
return outputs, logabsdet