Closed LanXiaoPang613 closed 2 weeks ago
besides, when i read the code, i found that the loss l_f_meta is zero when multiplied by eps. It indicates that the gradient is 0. So i want to understand the purpose of this operation.
Thanks for your interest in our work!
We follow previous implementation in dividemix to generate asymmetric noise.
For the initialization of l_f_meta, we follow the approach of Ren et al. in "Learning to Reweight Examples for Robust Deep Learning," where the initialization value is set to zero.
Hi Zhou, i want to know wheather the asymmetric label noise is generated like previous method unicon?Actually, I saw
flip
andflip2
in the code, but I did not see the asymmetric label noise generation function used inunicon
, such as cat<->dog, bird->plane, etc.