Open li-yiqing opened 1 year ago
I think the "or" was intentional to allow for cases of applying tv by setting only one of these variables (tv_before, tv_after) instead of both or for allowing tv before a certain point and after a certain point (discontinuous intervals [0, tv_before] + [tv_after, N_iters]). If you use "and", you can only apply TV for the iterations in the range [tv_after, tv_before].
Nice work. I am trying to read your code to implement it on paddle. but i notice something weird in
run.py
line 479, where you wroteif global_step<cfg_train.tv_before and global_step>cfg_train.tv_after and global_step%cfg_train.tv_every==0:
. i think it will not do tv regulation while trainingdid you mean
if global_step<cfg_train.tv_before or global_step>cfg_train.tv_after and global_step%cfg_train.tv_every==0:
, or i take this the wrong way``