5a126fc7cac272dfe2514554ef228001c795d729,slm_lab/agent/algorithm/sac.py,SoftActorCritic,act,#SoftActorCritic#Any#,85

Before Change

        if self.body.env.clock.frame < self.training_start_step:
            return policy_util.random(state, self, self.body).cpu().squeeze().numpy()
        else:
            action = super().act(state)
            return np.tanh(action)  // continuous action bound

    def calc_q(self, state, action, net=None):
        """Forward-pass to calculate the predicted state-action-value from q1_net."""

After Change

            return policy_util.random(state, self, self.body).cpu().squeeze().numpy()
        else:
            action = self.action_policy(state, self, self.body)
            if self.body.is_discrete:
                // discrete output is RelaxedOneHotCategorical, need to sample to int. clamp to prevent minor precision issue with prob < 0
                action = torch.distributions.Categorical(probs=action.clamp(min=0)).sample()
            else:
                action = torch.tanh(action)  // continuous action bound
            return action.cpu().squeeze().numpy()

    def calc_q(self, state, action, net=None):
        """Forward-pass to calculate the predicted state-action-value from q1_net."""
        x = torch.cat((state, action), dim=-1)

Instances