modified {left,right}-context determination code to take max of {init,ref}.config

[vijayaditya]

…ax of {init,ref}.config

[danpovey]

Looks like it's too late and you've already done it-- but since we spoke last, I realized that there is a problem with doing it this way. It causes a mismatch with the testing condition, where the extra-left-context would be determined from the model itself. This could impact results (in these unusual situations where there is this type of context mismatch). I thought of various code-level fixes and they were all quite complicated to implement. But I've thought of a different script-level fix that will re-use some of the code that you've already written - if you have time, I know you have a job now. Firstly, extend the lda-layer or whatever it's called by adding a "delay" parameter, defaulting to zero. If it's not zero, then when creating the init.config, instead of creating the output node directly, it will use that Descriptor that normally would have gone to the output node, and create an internal node (just call it delay-node). Then the output-node would be something like output-node name=output input=Offset(delay-node, t, $delay) If the script detects that the context of the init.config is not contained within the context of the ref.config, then it will die with an error message telling the user to add a "delay" option to the LDA layer. The script can actually work out the delay value that's closest to zero while also being sufficient to prevent the error from happening again, and can tell it to the user.

…

On Mon, Mar 27, 2017 at 8:22 PM, Vijayaditya Peddinti < ***@***.***> wrote: …ax of {init,ref}.config ------------------------------ You can view, comment on, or merge this pull request online at: #1516 Commit Summary - xconfig : modified {left,right}-context determination code to take max of {init,ref}.config File Changes - *M* egs/wsj/s5/steps/nnet3/xconfig_to_configs.py <https://github.com/kaldi-asr/kaldi/pull/1516/files#diff-0> (39) Patch Links: - https://github.com/kaldi-asr/kaldi/pull/1516.patch - https://github.com/kaldi-asr/kaldi/pull/1516.diff — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#1516>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADJVuye8ZMH0QWpAZJlWxXIz9FMm2LvMks5rqFLggaJpZM4Mq_RU> .

[danpovey]

should have mentioned- the internal node would just be a NoOpComponent.

…

On Mon, Mar 27, 2017 at 10:13 PM, Daniel Povey ***@***.***> wrote: Looks like it's too late and you've already done it-- but since we spoke last, I realized that there is a problem with doing it this way. It causes a mismatch with the testing condition, where the extra-left-context would be determined from the model itself. This could impact results (in these unusual situations where there is this type of context mismatch). I thought of various code-level fixes and they were all quite complicated to implement. But I've thought of a different script-level fix that will re-use some of the code that you've already written - if you have time, I know you have a job now. Firstly, extend the lda-layer or whatever it's called by adding a "delay" parameter, defaulting to zero. If it's not zero, then when creating the init.config, instead of creating the output node directly, it will use that Descriptor that normally would have gone to the output node, and create an internal node (just call it delay-node). Then the output-node would be something like output-node name=output input=Offset(delay-node, t, $delay) If the script detects that the context of the init.config is not contained within the context of the ref.config, then it will die with an error message telling the user to add a "delay" option to the LDA layer. The script can actually work out the delay value that's closest to zero while also being sufficient to prevent the error from happening again, and can tell it to the user. On Mon, Mar 27, 2017 at 8:22 PM, Vijayaditya Peddinti < ***@***.***> wrote: > …ax of {init,ref}.config > ------------------------------ > You can view, comment on, or merge this pull request online at: > > #1516 > Commit Summary > > - xconfig : modified {left,right}-context determination code to take > max of {init,ref}.config > > File Changes > > - *M* egs/wsj/s5/steps/nnet3/xconfig_to_configs.py > <https://github.com/kaldi-asr/kaldi/pull/1516/files#diff-0> (39) > > Patch Links: > > - https://github.com/kaldi-asr/kaldi/pull/1516.patch > - https://github.com/kaldi-asr/kaldi/pull/1516.diff > > — > You are receiving this because you are subscribed to this thread. > Reply to this email directly, view it on GitHub > <#1516>, or mute the thread > <https://github.com/notifications/unsubscribe-auth/ADJVuye8ZMH0QWpAZJlWxXIz9FMm2LvMks5rqFLggaJpZM4Mq_RU> > . >

[vijayaditya]

Seems simple enough. I will do this tomorrow.

…

--Vijay On Mon, Mar 27, 2017 at 7:14 PM, Daniel Povey <notifications@github.com> wrote:

Looks like it's too late and you've already done it-- but since we spoke last, I realized that there is a problem with doing it this way. It causes a mismatch with the testing condition, where the extra-left-context would be determined from the model itself. This could impact results (in these unusual situations where there is this type of context mismatch). I thought of various code-level fixes and they were all quite complicated to implement. But I've thought of a different script-level fix that will re-use some of the code that you've already written - if you have time, I know you have a job now. Firstly, extend the lda-layer or whatever it's called by adding a "delay" parameter, defaulting to zero. If it's not zero, then when creating the init.config, instead of creating the output node directly, it will use that Descriptor that normally would have gone to the output node, and create an internal node (just call it delay-node). Then the output-node would be something like output-node name=output input=Offset(delay-node, t, $delay) If the script detects that the context of the init.config is not contained within the context of the ref.config, then it will die with an error message telling the user to add a "delay" option to the LDA layer. The script can actually work out the delay value that's closest to zero while also being sufficient to prevent the error from happening again, and can tell it to the user. On Mon, Mar 27, 2017 at 8:22 PM, Vijayaditya Peddinti < ***@***.***> wrote: > …ax of {init,ref}.config > ------------------------------ > You can view, comment on, or merge this pull request online at: > > #1516 > Commit Summary > > - xconfig : modified {left,right}-context determination code to take > max of {init,ref}.config > > File Changes > > - *M* egs/wsj/s5/steps/nnet3/xconfig_to_configs.py > <https://github.com/kaldi-asr/kaldi/pull/1516/files#diff-0> (39) > > Patch Links: > > - https://github.com/kaldi-asr/kaldi/pull/1516.patch > - https://github.com/kaldi-asr/kaldi/pull/1516.diff > > — > You are receiving this because you are subscribed to this thread. > Reply to this email directly, view it on GitHub > <#1516>, or mute the thread > <https://github.com/notifications/unsubscribe-auth/ ADJVuye8ZMH0QWpAZJlWxXIz9FMm2LvMks5rqFLggaJpZM4Mq_RU> > . > — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#1516 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADtwoDxZjziW_annRS0JpEdqovg9zyLCks5rqGzrgaJpZM4Mq_RU> .

[vijayaditya]

Yep realized that. On Mon, Mar 27, 2017 at 7:15 PM, Daniel Povey <notifications@github.com> wrote:

…

should have mentioned- the internal node would just be a NoOpComponent. On Mon, Mar 27, 2017 at 10:13 PM, Daniel Povey ***@***.***> wrote: > Looks like it's too late and you've already done it-- but since we spoke > last, I realized that there is a problem with doing it this way. It causes > a mismatch with the testing condition, where the extra-left-context would > be determined from the model itself. This could impact results (in these > unusual situations where there is this type of context mismatch). > > I thought of various code-level fixes and they were all quite complicated > to implement. > > But I've thought of a different script-level fix that will re-use some of > the code that you've already written - if you have time, I know you have a > job now. > > Firstly, extend the lda-layer or whatever it's called by adding a "delay" > parameter, defaulting to zero. If it's not zero, then when creating the > init.config, instead of creating the output node directly, it will use that > Descriptor that normally would have gone to the output node, and create an > internal node (just call it delay-node). Then the output-node would be > something like > output-node name=output input=Offset(delay-node, t, $delay) > > If the script detects that the context of the init.config is not contained > within the context of the ref.config, then it will die with an error > message telling the user to add a "delay" option to the LDA layer. The > script can actually work out the delay value that's closest to zero while > also being sufficient to prevent the error from happening again, and can > tell it to the user. > > > > On Mon, Mar 27, 2017 at 8:22 PM, Vijayaditya Peddinti < > ***@***.***> wrote: > >> …ax of {init,ref}.config >> ------------------------------ >> You can view, comment on, or merge this pull request online at: >> >> #1516 >> Commit Summary >> >> - xconfig : modified {left,right}-context determination code to take >> max of {init,ref}.config >> >> File Changes >> >> - *M* egs/wsj/s5/steps/nnet3/xconfig_to_configs.py >> <https://github.com/kaldi-asr/kaldi/pull/1516/files#diff-0> (39) >> >> Patch Links: >> >> - https://github.com/kaldi-asr/kaldi/pull/1516.patch >> - https://github.com/kaldi-asr/kaldi/pull/1516.diff >> >> — >> You are receiving this because you are subscribed to this thread. >> Reply to this email directly, view it on GitHub >> <#1516>, or mute the thread >> <https://github.com/notifications/unsubscribe-auth/ ADJVuye8ZMH0QWpAZJlWxXIz9FMm2LvMks5rqFLggaJpZM4Mq_RU> >> . >> > > — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#1516 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADtwoLYKpgC7fBseKyaO6xZivzNWS4XAks5rqG0ogaJpZM4Mq_RU> .

[danpovey]

ping...

[vijayaditya]

@danpovey got busy with paper writing will do it by EoD tomorrow.

[danpovey]

ping..

[vijayaditya]

working on it .. should be done soon.

[vijayaditya]

See #1540 .