pix2seq
pix2seq copied to clipboard
Question about inference
During inference, the token (5-th) that is corresponded to the object class may be classified to the coordinates. In the other hand, the token that is corresponded to the coordinates still has chance to be classified to the class of the object. How to deal with such situation? Thanks a lot.
you can offset logits before sampling if you want to disable certain predictions (e.g. set -1e9 for logits that correspond to class/coordinates). but we find that just free sampling is fine.
Thanks for your kindly replying!