Question about the unused parameters in Adapointr

[pkqbajng]

Hi, sir, thanks for your excellent job. However, I found that the parameters of the query ranking won't be updated during the training procedure if set find_unused_paramters=False. Could you please check its effectiveness?

[yuxumin]

Yes, it wont be optimized. The right implementation of this ranking is based on gumble-softmax, we will correct the code for this part in next push.

[pkqbajng]

Okay, thanks for your reply. When conducting ablation study on this, does you concat the generated query with the input ones and then use farthest sampling in the original experiments?