Question about the weights in Softmatch

[Hugo-cell111]

Hi! I want to raise a question of SoftMatch.
When I try to adapt the Softmatch algorithm to my personal experiment, the weights will drop severly to the order of 1e-5, which means the weight mask doesn't work. I think the ideal range of the weight mask is 1e-1 to 1.
Have you conducted the experiment of the image-wise weights of loss function in this line of code? And what is the actual result? Thanks!