TGN: __assoc__ in LastNeighborLoader

[ssharpe42]

Is there a reason you don't flush self.__assoc__ after an insert or call and make it into torch.empty() again? Since you are always overwriting it anyways and it would eventually fill up over time would that be more memory efficient or is there a downside to that?

[rusty1s]

I don't think a flush is necessary, as we only need to temporally store the global to local node indices. Since __assoc__ is a vector that is stored in memory during the whole training procedure, and there is no memory overhead in simply updating its values. Let me know if that clarifies your doubts.

[ssharpe42]

Ok maybe I was just confused how torch.empty works. So after the whole training procedure when __assoc__ is filled with values, will it have the same memory requirement as it did when it was initialized with an empty tensor?

[rusty1s]

Yes. torch.empty will allocate memory just as every other tensor, it‘s just not pre-filled with any specific values.