67d13b4ba91b23ad29f660aae68a01ddbd809530,chainerrl/agents/ppo.py,PPO,act_and_train,#PPO#Any#Any#,407

Before Change

        self._update(dataset_iter, target_model, mean_advs, std_advs)

    def act_and_train(self, obs, reward):
        if hasattr(self.model, "obs_filter"):
            xp = self.xp
            b_state = self.batch_states([obs], xp, self.phi)
            self.model.obs_filter.experience(b_state)

        action, v = self._act(obs)

        // Update stats
        self.average_v += (

After Change

            b_state = self.obs_normalizer(b_state, update=False)

        // action_distrib will be recomputed when computing gradients
        with chainer.using_config("train", False), chainer.no_backprop_mode():
            action_distrib, value = self.model(b_state)
            action = chainer.cuda.to_cpu(action_distrib.sample().data)[0]
            self.entropy_record.append(float(action_distrib.entropy.data))

Instances