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
Hello, I'm trying to understand the behaviour of the individual losses and when I look at the implementation, I see that the loss_dict is what is logging different losses for dino local crops, global crops, ibot loss and koleo loss.
In the loss dict, here the ibot loss is divided by 2. But when calculatin the gradients and when the backward call happens, the ibot loss isn't divided by 2 here and then ultimately in the total_losshere metric too, this loss is not divided by 2.
So why is this loss during metric logging divided by 2? Is this just a mistake in logging or the actual ibot loss should be divided by 2?
The text was updated successfully, but these errors were encountered:
Hello, I'm trying to understand the behaviour of the individual losses and when I look at the implementation, I see that the
loss_dict
is what is logging different losses for dino local crops, global crops, ibot loss and koleo loss.In the loss dict, here the ibot loss is divided by 2. But when calculatin the gradients and when the
backward
call happens, the ibot loss isn't divided by 2 here and then ultimately in thetotal_loss
here metric too, this loss is not divided by 2.So why is this loss during metric logging divided by 2? Is this just a mistake in logging or the actual ibot loss should be divided by 2?
The text was updated successfully, but these errors were encountered: