Closed DISAPPEARED13 closed 1 year ago
That's purely empirical. You may check out the ablative analysis section to see the performance of other sample numbers.
Thanks for replying! I am confused that the sample numbers should have been related with the input patch pixels number. For enough points seems play a very important role when caculating the reco loss :smile:
So is it the reason when you chose number of queries
and number of keys
?
Strictly speaking yes. But it works well in practise which I assume mainly because of the design of active sampling.
Thanks for inspiring work!
I notice that in part Implementation Details , you mention that in ReCo framework, you sampled 256 query samples and 512 key samples for each mini-batch, could you please tell me why is 256 and 512?
Thanks a lot!