suruoxi / distanceweightedsampling Goto Github PK

Hello. I am getting a weird effect where an epoch or so into training, I will get nan's as part of the get_distance output (and thus nan's for all the weights). In the process of debugging but I'm wondering if you ever encountered this in training? Thanks.

EDIT: Fixed. Not sure why, but occasionally the line 2-2*sim will have small negative values, which produces nan with the sqrt. Masking out small values to 0 fixes the problem.

First of all, thank you very much for coding as Pytorch.

I have a question in that code.

I have a question about the expression for getting log_weights in the Distance Weighted Sampling Class.

First, in this paper, the expression is written as follows:

The current code is as follows.
log_weights = ((2.0 - float(d)) * distance.log() - (float(d-3)/2)torch.log(torch.clamp(1.0 - 0.25(distance*distance), min=1e-8)))

Is this code correct to code the formula for giving weight?: