Selection issues

[gotobcn8]

            if self.server_solver.do_selection:
                for s in range(self.num_clusters):
                    selection.append(np.random.choice(a=range(self.num_clients), size=self.server_solver.selection_size,
                                                      p=self.importance_weights_matrix[:, s], replace=False).tolist())
                logger.log_client_selection(self.exp_id, t, self._idx_to_id(selection))
            else:
                selection = np.tile(range(self.num_clients), reps=(self.num_clusters, 1))

            # Local updates
            for k in np.unique(np.concatenate(selection).ravel()):
                self.client_vec[k].run()

In this code, if server_solver.do_selection, you will select repeatitive clients right? and following that you use unique() to remove the repeaetitive one. It's possibly selection >= server_solver.selection_size.
But the question is when you aggregate the cluster. why you just do the repetitive operations for each cluster?
Plz see my notations.

def _aggregate_fedsoft(self, selection):
        for s in range(self.num_clusters):
            next_weights = self.generate_zero_weights()
            for k in selection[s]:
                if self.server_solver.do_selection:
                    # why the weight is 1 / server_solver.selection_size not len(selection)?
                    aggregation_weight = 1. / self.server_solver.selection_size
                else:
                    aggregation_weight = self.importance_weights_matrix[k][s]
                client_weights = self.client_vec[k].get_model_dict()
                for key in next_weights.keys():
                    # why you just do aggregation for selection in cluster_num times?
                    next_weights[key] += aggregation_weight * client_weights[key].cpu()
            self.cluster_vec[s].load_state_dict(state_dict=next_weights)

After I run the code, it really confused me cuz not as my expectation.
Thanks