About gating_top_n

[Heihaierr]

Hi, I notice there is experiment with top_n=1 in the paper of st-moe. But in st_moe_pytorch.py,
assert top_n >= 2, 'must be 2 or more experts'
Can top_n=1 work in this implementation?

[lucidrains]

@Heihaierr not here, as it was just to be faithful to the paper, which explored 2 and then a generalization of top-n (iirc) up to 3 and 4

i thought that top 1 didn't work that well?

[Heihaierr]

@Heihaierr not here, as it was just to be faithful to the paper, which explored 2 and then a generalization of top-n (iirc) up to 3 and 4

i thought that top 1 didn't work that well?

Yes, but the paper also explored top-1 routing and shows improvement.

[lucidrains]

ah I see, yeah they did, but 2 is still recommended

[Heihaierr]

get it, thanks for quick reply

[moon4869]

If top_n=1, how should we achieve it?