search

kaldi-asr / kaldi

kaldi-asr/kaldi is the official location of the Kaldi project.
http://kaldi-asr.org
Other
14.24k stars 5.32k forks source link

Kaldi nnet2 online training with ivectors #2784

Closed Umar17 closed 5 years ago

Umar17 commented 6 years ago

Hello all,

I am trying to build my own nnet2 model for online decoding through kaldi-gstreamer that requires ivectors as input feature. I have tried _train_tanhfast.sh but it fails with KALDI ASSERTION on 'tot_dim == nnet.InputDim()'. Can anyone please guide me how to build kaldi online model with ivectors (which script should I use? tanh, pnorm or whichever?)

Best Regards

danpovey commented 6 years ago

pnorm should be better than tanh. However the nnet2 scripts are outdated and I don't remember much about them. I suggest to use the nnet3 scripts; the mini_librispeech setup is a good example. In general rather than trying to figure out how to use a script yourself, it's better to search for examples of scripts that invoke it, and copy-and-modify that.

Dan

On Thu, Oct 11, 2018 at 7:42 AM Umar17 notifications@github.com wrote:

Hello all,

I am trying to build my own nnet2 model for online decoding through kaldi-gstreamer that requires ivectors as input feature. I have tried train_tanh_fast.sh but it fails with KALDI ASSERTION on 'tot_dim == nnet.InputDim()'. Can anyone please guide me how to build kaldi online model with ivectors (which script should I use? tanh, pnorm or whichever?)

Best Regards

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/kaldi-asr/kaldi/issues/2784, or mute the thread https://github.com/notifications/unsubscribe-auth/ADJVu_-ycJun5ak72HSOQsYKy7pL3JI8ks5ujy6TgaJpZM4XXXDq .

313024351 commented 5 years ago

pnorm should be better than tanh. However the nnet2 scripts are outdated and I don't remember much about them. I suggest to use the nnet3 scripts; the mini_librispeech setup is a good example. In general rather than trying to figure out how to use a script yourself, it's better to search for examples of scripts that invoke it, and copy-and-modify that. Dan … On Thu, Oct 11, 2018 at 7:42 AM Umar17 @.**> wrote: Hello all, I am trying to build my own nnet2 model for online decoding through kaldi-gstreamer that requires ivectors as input feature. I have tried train_tanh_fast.sh* but it fails with KALDI ASSERTION on 'totdim == nnet.InputDim()'. Can anyone please guide me how to build kaldi online model with ivectors (which script should I use? tanh, pnorm or whichever?) Best Regards — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#2784>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADJVu-ycJun5ak72HSOQsYKy7pL3JI8ks5ujy6TgaJpZM4XXXDq> .

the same question occured when i using nnet2/train_convnet_accel2.sh with online_ivector_dir. and if no online_ivector_dir, it will be ok.

danpovey commented 5 years ago

You should be using more up-to-date scripts. Search for examples called local/nnet3/run_tdnn.sh.

On Sun, Dec 16, 2018 at 7:56 PM 韩佳俊 notifications@github.com wrote:

pnorm should be better than tanh. However the nnet2 scripts are outdated and I don't remember much about them. I suggest to use the nnet3 scripts; the mini_librispeech setup is a good example. In general rather than trying to figure out how to use a script yourself, it's better to search for examples of scripts that invoke it, and copy-and-modify that. Dan … <#m22687386392257608> On Thu, Oct 11, 2018 at 7:42 AM Umar17 @.**> wrote: Hello all, I am trying to build my own nnet2 model for online decoding through kaldi-gstreamer that requires ivectors as input feature. I have tried train_tanh_fast.sh* but it fails with KALDI ASSERTION on 'totdim == nnet.InputDim()'. Can anyone please guide me how to build kaldi online model with ivectors (which script should I use? tanh, pnorm or whichever?) Best Regards — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#2784 https://github.com/kaldi-asr/kaldi/issues/2784>, or mute the thread https://github.com/notifications/unsubscribe-auth/ADJVu-ycJun5ak72HSOQsYKy7pL3JI8ks5ujy6TgaJpZM4XXXDq .

the same question occured when i using nnet2/train_convnet_accel2.sh with online_ivector_dir. and if no online_ivector_dir, it will be ok.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/kaldi-asr/kaldi/issues/2784#issuecomment-447715544, or mute the thread https://github.com/notifications/unsubscribe-auth/ADJVu_m9JTjbAWSBt3n2A4n9lyf1w0PXks5u5xXRgaJpZM4XXXDq .

Umar17 commented 5 years ago

Yeah I was able to build a tdnn model with ivectors using the same script (local/nnet3/run_tdnn.sh)

danpovey commented 5 years ago

OK. And sorry, I mistyped... the recommended scripts would normally be called local/chain/run_tdnn.sh these will give better WERs.

On Sun, Dec 16, 2018 at 10:52 PM Umar17 notifications@github.com wrote:

Yeah I was able to build a tdnn model with ivectors using the same script (local/nnet3/run_tdnn.sh)

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/kaldi-asr/kaldi/issues/2784#issuecomment-447740691, or mute the thread https://github.com/notifications/unsubscribe-auth/ADJVu7I_mJxGyckcMS2s2Mq8dJq4VTVyks5u5z8ygaJpZM4XXXDq .