You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
This repository has been archived by the owner on Jul 20, 2022. It is now read-only.
I notice that if no_weight_decay_on_bn is set to True, weight decay will only apply to conv.weight. It seems that weight decay on fc layers are also removed at the same time. Is there any reason to do so?
The text was updated successfully, but these errors were encountered:
I think I added the option to try what's written in section 4.2 of this paper, and if I remember correctly, in my experiments, the accuracy was higher when no weight decay was applied to the fc layer. So that's the reason.
But the accuracy drop may be because I used the same parameter for weight decay on conv and fc layers, and tuning weight decay parameter on fc layer could lead to better accuracy. So in that sense, the current implementation that forces the weight decay on the fc layer to be 0 is not very good.
I notice that if
no_weight_decay_on_bn
is set toTrue
, weight decay will only apply toconv.weight
. It seems that weight decay on fc layers are also removed at the same time. Is there any reason to do so?The text was updated successfully, but these errors were encountered: