docstring in DecoupledAdaLRLion is not cohere with the code
ericxsun opened this issue · comments
QinLuo commented
The docstring of DecoupledAdaLRLion say the LR is scaled down by min(`lr_penalty` ** N, `min_scale`)
, but the code implemented adjust_lr
here is lr * max(min_scale, lr_penalty**num_times)
.
So which is the right one, max
or min
?