def _build(self, inputs):
mean, covariance, scale = self.create_mean_n_cov_layers(inputs)
mean_plus_minus_one = self.force_between_plus_minus_one(
mean, self.plus_minus_one_method)
# TODO if you want contractive regularizers implement them first. Then, uncomment the following lines (Riccardo)
# self.set_contractive_regularizer(mean_zero_one, covariance,
# self._contractive_regularizer_inputs,
# self._contractive_regularizer_tuple,
# self._contractive_collection_network_str)
#
output_distribution = tfd.Logistic(loc=mean_plus_minus_one,
scale=scale)
# add reconstruction_node method (needed to some sort of mean or median to get reconstructions without sampling)
def reconstruction_node(self):
return self.mean()
output_distribution.reconstruction_node = types.MethodType(
reconstruction_node, output_distribution)
return output_distribution
def discrete_logistics_likelihood(loc, scale, obs, side_censored=True):
from tensorflow_probability import distributions as tfd
x_lo = obs - 1 / 255 / 2
x_hi = obs + 1 / 255 / 2
if side_censored:
x_lo = tf.where(x_lo >= 0, x_lo,
tf.to_float(-1000) * tf.ones_like(x_lo))
x_hi = tf.where(x_hi <= 1, x_hi,
tf.to_float(+1000) * tf.ones_like(x_hi))
dist = tfd.Logistic(tf.to_double(loc), tf.to_double(scale))
# tf.where doesn't backpropagate correctly when there's NaN in the unused branch. So instead of
# ret = tf.where(
# tf.math.abs(obs - loc) <= 2*scale,
# tf.math.log(dist.cdf(tf.to_double(x_hi)) - dist.cdf(tf.to_double(x_lo))),
# dist.log_prob(tf.to_double(obs)) + tf.to_double(tf.math.log(1/255)))
# we need
p_exact = dist.cdf(tf.to_double(x_hi)) - dist.cdf(tf.to_double(x_lo))
p_exact = tf.where(
tf.math.abs(obs - loc) <= 2 * scale, p_exact, tf.ones_like(p_exact))
ret = tf.where(
tf.math.abs(obs - loc) <= 2 * scale, tf.math.log(p_exact),
dist.log_prob(tf.to_double(obs)) + tf.to_double(tf.math.log(1 / 255)))
return tf.to_float(ret)
def _init_distribution(conditions, **kwargs):
loc, scale = conditions["loc"], conditions["scale"]
return tfd.Logistic(loc=loc, scale=scale, **kwargs)
def _base_dist(self, mu: TensorLike, s: TensorLike, *args, **kwargs):
return tfd.Logistic(loc=mu, scale=s, *args, **kwargs)