9bdbb11cf27060e7847a87dcdf691dd6b96ce6df,rlkit/data_management/obs_dict_replay_buffer.py,ObsDictRelabelingBuffer,random_batch,#ObsDictRelabelingBuffer#,171

Before Change

                // This is generally faster than random.choice. Makes you wonder what
                // random.choice is doing
                num_options = len(possible_future_obs_idxs)
                next_obs_i = int(np.random.randint(0, num_options))
                future_obs_idxs.append(possible_future_obs_idxs[next_obs_i])
            future_obs_idxs = np.array(future_obs_idxs)
            resampled_goals[-num_future_goals:] = self._next_obs[
                self.achieved_goal_key

After Change

            //// better future obs sample algorithm
            future_indices = indices[-num_future_goals:]
            possible_future_obs_lens = np.array([len(self._idx_to_future_obs_idx[i]) for i in future_indices])
            next_obs_idxs = (np.random.random(num_future_goals) * possible_future_obs_lens).astype(np.int)
            future_obs_idxs = np.array([self._idx_to_future_obs_idx[ids][next_obs_idxs[i]] for i, ids in enumerate(future_indices)])

            resampled_goals[-num_future_goals:] = self._next_obs[
                self.achieved_goal_key
            ][future_obs_idxs]

Instances