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

Before Change

        // Update stats
        self.average_v += (
            (1 - self.average_v_decay) *
            (v[0] - self.average_v))

        if self.last_state is not None:
            self.last_episode.append({

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))
            self.value_record.append(float(value.data))

        if self.last_state is not None:
            self.last_episode.append({
                "state": self.last_state,
                "action": self.last_action,

Instances