search

Dannynis / xvector_pytorch

A pytorch implementation of xvector embedding
78 stars 8 forks source link

loss change to nan #6

Closed HappyGallery closed 3 years ago

HappyGallery commented 4 years ago

hey, first thanks for the great help of your code 。I using the dataset called cnceleb ,but i have faced a problem .As the training goes on the loss change to nan ,the features i feed in the net is mfcc features gained from libROSA . And i have been debugged it for about 2 weeks. input = (64,24,126) (batchsize,feature_dim,num_of_mfcc )

lengjiayi commented 4 years ago

Hi, I used to have the same issue: loss change to nan after first iteration. Finally I fix it by simply remove dropout in tdnn. Hope this might help you :)

HappyGallery commented 3 years ago

Thanks for your advice, i would like to have a try