9527891f31435622726d8bd1fa84defe6fa29dbe,torch_ac/algos/ppo.py,PPOAlgo,update_parameters,#PPOAlgo#,22

Before Change

                b.old_log_prob = b.old_log_prob.detach()
                b.advantage = b.advantage.detach()
                b.old_value = b.old_value.detach()
                b.returnn = b.returnn.detach()

                // Compute loss

                preprocessed_obs = self.preprocess_obss(b.obs, requires_grad=True, use_gpu=torch.cuda.is_available())

After Change

        // Add old action log probs and old values to transitions

        preprocessed_obs = self.preprocess_obss(ts.obs, use_gpu=torch.cuda.is_available())
        with torch.no_grad():
            ts.old_log_prob = self.acmodel.get_dist(preprocessed_obs).log_prob(ts.action)
            ts.old_value = self.acmodel.get_value(preprocessed_obs)

        if self.batch_size == 0:
            self.batch_size = len(ts)

        for _ in range(self.epochs):

Instances