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
It seems the speed_loss is being divided by params['branches'][0].shape[0] (the batch_size?) twice instead of only once. While the rest of the loss ('loss_function') is not.
Is this indeed a bug that changes the scaling of the different losses or am I missing something?
The text was updated successfully, but these errors were encountered:
Yes, I found this recently. It was dividing twice.
Thanks for pointing this out. I wouldnt' change that on the repository since this is necessary
in order to reproduce the results.
This shows that the speed prediction should probably use a way smaller
weight than what i used originally.
In the file
coiltraine/network/loss.py
we find the following lines from 56 onwards:
It seems the speed_loss is being divided by
params['branches'][0].shape[0]
(the batch_size?) twice instead of only once. While the rest of the loss ('loss_function') is not.Is this indeed a bug that changes the scaling of the different losses or am I missing something?
The text was updated successfully, but these errors were encountered: