910e4f89faf44e46c06eecc80aab3dff3ca1c5dc,softlearning/policies/gaussian_policy.py,GaussianPolicy,_squash_correction,#GaussianPolicy#Any#,122

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):
        Context manager for changing the determinism of the policy.

After Change

    def _squash_correction(self, actions):
        if not self._squash: return 0
        // more stable squash correction
        return tf.reduce_sum(2. * (tf.log(2.) - actions - tf.nn.softplus(-2. * actions)))
        // return tf.reduce_sum(tf.log(1 - tf.tanh(actions) ** 2 + EPS), axis=1)

    @contextmanager
    def deterministic(self, set_deterministic=True):

Instances