algorithm_badmm.py 文件源码

def _set_interp_values(self):
        """
        Use iteration-based interpolation to set values of some
        schedule-based parameters.
        """
        # Compute temporal interpolation value.
        t = min((self.iteration_count + 1.0) /
                (self._hyperparams['iterations'] - 1), 1)
        # Perform iteration-based interpolation of entropy penalty.
        if type(self._hyperparams['ent_reg_schedule']) in (int, float):
            self.policy_opt.set_ent_reg(self._hyperparams['ent_reg_schedule'])
        else:
            sch = self._hyperparams['ent_reg_schedule']
            self.policy_opt.set_ent_reg(
                np.exp(np.interp(t, np.linspace(0, 1, num=len(sch)),
                                 np.log(sch)))
            )
        # Perform iteration-based interpolation of Lagrange multiplier.
        if type(self._hyperparams['lg_step_schedule']) in (int, float):
            self._hyperparams['lg_step'] = self._hyperparams['lg_step_schedule']
        else:
            sch = self._hyperparams['lg_step_schedule']
            self._hyperparams['lg_step'] = np.exp(
                np.interp(t, np.linspace(0, 1, num=len(sch)), np.log(sch))
            )