search

google / uis-rnn

This is the library for the Unbounded Interleaved-State Recurrent Neural Network (UIS-RNN) algorithm, corresponding to the paper Fully Supervised Speaker Diarization.
https://arxiv.org/abs/1810.04719
Apache License 2.0
1.55k stars 320 forks source link

[Question] About num_non_zero #84

Closed shubhunegi30 closed 3 years ago

shubhunegi30 commented 3 years ago

I am trying to train UISRNN on VoxConverse dataset. But from the very first iteration, Sigma Prior Loss is coming as Nan which is affecting other losses too and after first iteration every loss is becoming Nan due to addition with Sigma Prior Loss.

I checked the values of tensors while calculating the particular loss and found that some of the values of the variable 'num_non_zero' are zeroes. And due to the variable being in denominator in the loss calculation, it's becoming inf. That's the reason for Nan values. What should I do? What is possibly causing this?

wq2012 commented 3 years ago

Did you check your data?

Likely it's due to you are passing zero-length sequences as input.