Questions regarding the cls loss

[leonard-git-acc]

Hello,
I have been looking through your code for the cls loss and I have come across two places, where I have some questions.
The first place is here. I am not quite sure why the sum of the probabilities is clamped. Is this better, than, for example, normalizing it?

probs_ = torch.clamp(probs_, 0., 1.)

My second question concerns the same file in this location. Here there is the variable data_balancing_weight declared as 0 and never assigned a new value. It is then multiplied with occ_pts_prob_error.mean(), which then, if I am not missing anything, should not have any impact in the loss calculation. Is this intended?

data_balancing_weight=0
occ_pts_prob_error = torch.abs(1 - occ_pts_prob)
free_pts_prob_error = torch.abs(0 - free_pts_prob)
            
# L1 LOSS
cls_loss = occ_pts_prob_error.mean()*data_balancing_weight + free_pts_prob_error.mean()
loss += cls_loss

Thank you for your help!

[kungfrank]

Hello,

Thanks for your questions!

1. I clamp the queried probability rather than normalize it because I wanted to preserve the occupied probability at the (x, y, z) position, which carries physical significance. Normalizing it would distort this meaning.

2. You're right. The data_balancing_weight was used in a previous experiment, but I found it unhelpful for my current method, so I have disabled it.