def __call__(self, inputs, *args, **kwargs):
if not self.built:
mean, variance = tf.nn.moments(
inputs, axes=list(range(inputs.shape.ndims - 1)))
self.bias_initial_value = -mean
# TODO(trandustin): Optionally, actnorm multiplies log_scale by a fixed
# log_scale factor (e.g., 3.) and initializes by
# initial_value / log_scale_factor.
self.log_scale_initial_value = tf.math.log(
1. / (tf.sqrt(variance) + self.epsilon))
if not isinstance(inputs, random_variable.RandomVariable):
return super(ActNorm, self).__call__(inputs, *args, **kwargs)
return transformed_random_variable.TransformedRandomVariable(inputs, self)
def __call__(self, inputs, *args, **kwargs):
if not isinstance(inputs, random_variable.RandomVariable):
return super(SinkhornAutoregressiveFlow,
self).__call__(inputs, *args, **kwargs)
return transformed_random_variable.TransformedRandomVariable(
inputs, self)
def __call__(self, inputs, *args, **kwargs):
if not isinstance(inputs, random_variable.RandomVariable):
return super(DiscreteBipartiteFlow,
self).__call__(inputs, *args, **kwargs)
return transformed_random_variable.TransformedRandomVariable(
inputs, self)
