Method spatial_att_feed_forward_generator

storage.py:205–281

(self, advantages)

                  temporal_hidden_states

    def spatial_att_feed_forward_generator(self, advantages):
        # mini_batch_size = (self.num_steps- params.rnn_seq_len) // self.mini_batch_num  # 500-5/4
        mini_batch_size = params.batch_size
        sampler = BatchSampler(SubsetRandomSampler(range(self.num_steps - params.rnn_seq_len)), mini_batch_size,
                               drop_last=True)
        T = params.rnn_seq_len
        N = mini_batch_size
        for indices in sampler:
            obs_batch = self.obs[indices]
            uav_aoi_batch = self.uav_aoi[indices]
            uav_snr_batch = self.uav_snr[indices]
            uav_tuse_batch = self.uav_tuse[indices]
            uav_effort_batch = self.uav_effort[indices]
            action_dia_batch = self.action_dia[indices]
            value_pred_batch = self.value_preds[indices]
            return_batch = self.returns[indices]
            old_action_log_probs_batch = self.action_log_probs[indices]
            advantages_batch = advantages[indices]
            masks_batch = self.masks[indices]
            temporal_hidden_states = self.temporal_hidden_states[indices]
            spatial_hidden_states=self.spatial_hidden_states[indices]
            sample_index = np.array(indices)
            for t in range(1, params.rnn_seq_len):
                cur_index = sample_index + t
                cur_index = cur_index.tolist()
                obs_batch_i = self.obs[cur_index]
                uav_aoi_batch_i = self.uav_aoi[cur_index]
                uav_snr_batch_i = self.uav_snr[cur_index]
                uav_tuse_batch_i = self.uav_tuse[cur_index]
                uav_effort_batch_i = self.uav_effort[cur_index]
                action_dia_batch_i = self.action_dia[cur_index]
                value_pred_batch_i = self.value_preds[cur_index]
                return_batch_i = self.returns[cur_index]
                old_action_log_probs_batch_i = self.action_log_probs[cur_index]
                advantages_batch_i = advantages[cur_index]
                masks_batch_i = self.masks[cur_index]
                if t == 1:
                    obs_batch = torch.stack((obs_batch, obs_batch_i))
                    uav_aoi_batch = torch.stack((uav_aoi_batch, uav_aoi_batch_i))
                    uav_snr_batch = torch.stack((uav_snr_batch, uav_snr_batch_i))
                    uav_tuse_batch = torch.stack((uav_tuse_batch, uav_tuse_batch_i))
                    uav_effort_batch = torch.stack((uav_effort_batch, uav_effort_batch_i))
                    action_dia_batch = torch.stack((action_dia_batch, action_dia_batch_i))
                    value_pred_batch = torch.stack((value_pred_batch, value_pred_batch_i))
                    return_batch = torch.stack((return_batch, return_batch_i))
                    old_action_log_probs_batch = torch.stack((old_action_log_probs_batch, old_action_log_probs_batch_i))
                    advantages_batch = torch.stack((advantages_batch, advantages_batch_i))
                    masks_batch = torch.stack((masks_batch, masks_batch_i))
                else:
                    obs_batch = torch.cat((obs_batch, obs_batch_i.unsqueeze(0)), dim=0)
                    uav_aoi_batch = torch.cat((uav_aoi_batch, uav_aoi_batch_i.unsqueeze(0)), dim=0)
                    uav_snr_batch = torch.cat((uav_snr_batch, uav_snr_batch_i.unsqueeze(0)), dim=0)
                    uav_tuse_batch = torch.cat((uav_tuse_batch, uav_tuse_batch_i.unsqueeze(0)), dim=0)
                    uav_effort_batch = torch.cat((uav_effort_batch, uav_effort_batch_i.unsqueeze(0)), dim=0)
                    action_dia_batch = torch.cat((action_dia_batch, action_dia_batch_i.unsqueeze(0)), dim=0)
                    value_pred_batch = torch.cat((value_pred_batch, value_pred_batch_i.unsqueeze(0)), dim=0)
                    return_batch = torch.cat((return_batch, return_batch_i.unsqueeze(0)), dim=0)
                    old_action_log_probs_batch = torch.cat(