dcc34ef8cd1e55fee69eb9f764650fdbab9da18f,softlearning/policies/gmm.py,GMMPolicy,_squash_correction,#GMMPolicy#Any#,159

Before Change

    def _squash_correction(self, actions):
        if not self._squash: return 0
        return tf.reduce_sum(tf.log(1 - tf.tanh(actions) ** 2 + EPS), axis=1)

    @contextmanager
    def deterministic(self, set_deterministic=True, latent=None):
        Context manager for changing the determinism of the policy.

After Change

        // return tf.reduce_sum(tf.log(1 - tf.tanh(actions) **2 + EPS), axis=1)

        // numerically stable squash correction without bias from EPS
        return tf.reduce_sum(2. * (tf.log(2.) - actions - tf.nn.softplus(-2. * actions)), axis=1)

    @contextmanager
    def deterministic(self, set_deterministic=True, latent=None):
        Context manager for changing the determinism of the policy.

Instances