49c887e9208d2a0119eba2a26e38d033bb2ba4da,sac/policies/real_nvp.py,RealNVPPolicy,__init__,#RealNVPPolicy#Any#Any#Any#Any#,18

Before Change

        self.distribution = RealNVP(real_nvp_config)

        super().__init__(
            env_spec,
            obs_pl=self._observations_ph,
            action=(
                tf.tanh(self._distribution_z)
                if squash
                else self._distribution_z
            ),
            scope_name="policy"
        )

    @overrides
    def get_action(self, observations):
        Sample action based on the observations.

After Change

        ds = tf.contrib.distributions
        self.bijector = RealNVPBijector(config=real_nvp_config,
                                        event_ndims=self._Ds)
        self.base_distribution = ds.MultivariateNormalDiag(
            loc=tf.zeros(self._Ds), scale_diag=tf.ones(self._Ds))

        self.distribution = ds.TransformedDistribution(
            distribution=self.base_distribution,
            bijector=self.bijector,
            name="RealNVPPolicyDistribution")


        y = self.distribution.bijector.forward(self._observations_ph)
        self.log_pi = self.distribution.log_prob(y)
        self.pi = tf.exp(self.log_pi)
        self._action = tf.tanh(self.pi) if squash else self.pi

    def get_action(self, observations):
        Sample action based on the observations.

Instances