Inconsistent Results Across Runs

[luyi256]

Hi! Thank you for providing such an efficient framework. I’ve encountered an issue where running the same code twice produces different outcomes, even after using seed_everything with the same seed and setting loc_seed in sample_core.cpp as a constant. I found that this inconsistency is caused by two main factors:

1. The update_memory function of class MailBox may update the same indices of a tensor simutaneously (Specifically, the line self.node_memory[nid.long()] = memory and self.node_memory_ts[nid.long()] = ts) since nid may contain duplicate IDs.
2. The update_mailbox function of class MailBox attempts to handle the issue mentioned above. However, the implementation does not work as expected. The line perm = inv.new_empty(uni.size(0)).scatter_(0, inv, perm) will not return the same perm each time for the same reason as the first factor.

To resolve these issues and obtain consistent results across runs (in addition to using seed_everything and setting loc_seed), I made the following modifications.
For theupdate_memory function, I added:

np_nid = nid.detach().cpu().numpy()
reversed_indices = np_nid[::-1]
unique_indices, first_indices = np.unique(
    reversed_indices, return_index=True)
last_indices = len(nid)-first_indices-1
tc_last_indices = torch.from_numpy(
    last_indices).to(self.device)
nid = nid[tc_last_indices]
memory = memory[tc_last_indices]
ts = ts[tc_last_indices]

before

self.node_memory[nid.long()] = memory
self.node_memory_ts[nid.long()] = ts

Similar modifications were applied to update_mailbox. With the modifications, I obtained the same outcomes for two runs.
I hope this helps clarify the problem and look forward to improved solutions from you. Thank you for your attention to this matter.