def get_expected_value(
self,
*inputs,
**kwargs
):
return tf.math.tanh(Gaussian.get_expected_value(
self, *inputs, **kwargs))
def get_log_probs(
self,
x,
*inputs,
**kwargs
):
x = tf.clip_by_value(x, -0.999, 0.999)
correction = -1.0 * tf.reduce_sum(tf.math.log(1.0 - tf.math.square(x)), axis=-1)
return correction + Gaussian.get_log_probs(
self, tf.math.atanh(x), *inputs, **kwargs)
def get_expected_value_of_prior(
self,
shape,
**kwargs
):
return tf.math.tanh(Gaussian.get_expected_value_of_prior(self, shape, **kwargs))
def sample_from_prior(
self,
shape,
**kwargs
):
return tf.math.tanh(Gaussian.sample_from_prior(self, shape, **kwargs))
def sample(
self,
*inputs,
**kwargs
):
return tf.math.tanh(Gaussian.sample(self, *inputs, **kwargs))
def get_log_probs(self, x, *inputs, **kwargs):
x = tf.maximum(x, 0.001)
correction = -1.0 * tf.reduce_sum(x, axis=-1)
return correction + Gaussian.get_log_probs(self, tf.math.log(x), *
inputs, **kwargs)