search

SeanNaren / warp-ctc

Pytorch Bindings for warp-ctc
Apache License 2.0
756 stars 271 forks source link

CTC loss is 0 on GPU Tesla V100 #162

Open bjlgcxc opened 4 years ago

bjlgcxc commented 4 years ago

I get the problem of continuously get zero loss (loss = 0) on GPU Tesla V100，if anyone got the same problem or have some solutions?

raotnameh commented 4 years ago

facing the same problem on longer audios of around 40-60 seconds. @SeanNaren

bjlgcxc commented 4 years ago

I have solved the problem, just put the input of ctc loss on cpu

for example , CTC(acts, labels, act_lens, label_lens),

do acts.cpu() before run CTC loss

Stonesjtu commented 4 years ago

The ctc loss will return 0 on errors. One is the ctc loss is not compiled with cuda support but got cuda tensor input. Another situation is input sequence is too long.

IAASSIBLCU commented 2 years ago

The ctc loss will return 0 on errors. One is the ctc loss is not compiled with cuda support but got cuda tensor input. Another situation is input sequence is too long.

I also face the same problem. so I want to know how to check the ctc loss is compiled with cuda or not. and, if it really isn't compiled, what can I do to make them compiled. thanks a lot!!!

raotnameh commented 2 years ago

@IAASSIBLCU I would recommend you to try using PyTorch CTC.

SeanNaren commented 2 years ago

I too would also suggest using PyTorch CTC :) this library is fairly old, and I've also moved this project to use it successfully: https://github.com/SeanNaren/deepspeech.pytorch

IAASSIBLCU commented 2 years ago

@raotnameh @SeanNaren thanks a lot!! I will try PyTorch CTC in my project. Hope to get the dream result.