Segmentation fault (core dumped)

[Shylock]

Thank you for open-sourcing such a great project. We have successfully installed pkwrap，but when import this tool, it will throws the following question:
"
Type "help", "copyright", "credits" or "license" for more information.

import pkwrap
Segmentation fault (core dumped)
"
And we use the following environment:
python 3.7.7
pytorch 1.7.0+cuda10.1
kaldi 2020 edition
gcc and g++ are all 5.4.0 edition

Looking forward to your reply
shylock

[Srikanth M R]

Hello,

Are you trying to run any particular script, or are you just trying to import it in a Python shell?

Srikanth

[Shylock]

Dear Srikanth，
We just trying to import Pkwrapin a Python shell.
And we successfully installed pkwrap according to your prompt, but could not import Pkwrap.
Shylock

[Srikanth M R]

I see. Then please try importing torch first. In fact, try running

python test/test-import.py

to see if the test passes.

Other issues could be:

1. Kaldi library is not in your $LD_LIBRARY_PATH. See here
2. There is another library in your environment that is clashing with the Kaldi you used to compile pkwrap.

Let me know if that helps.

[Shylock]

Thanks for your kind reply.
We first set Kaldi library in $LD_LIBRARY_PATH.
My torch version is '1.7.0+cu101'.
When we trying th run 'python test/test-import.py', it will also throw the following error.
"Segmentation fault (core dumped)"

I dont know how to solve this problem? and i also want to use lf-smbr loss function to train ASR acoustic model based on Pkwrap.

Shylock

[Shylock]

We set Kaldi library just by the following code.
"export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$KALDI_ROOT/src/lib/"

And we not use your method "Kaldi library is not in your $LD_LIBRARY_PATH. See [here]".

[Srikanth M R]

Could you please share the version of Kaldi (the output you get from git log)?

And which recipe/script in Kaldi runs lf-smbr loss? I can try to wrap it here. Currently, only the vanilla LF-MMI (both hybrid and E2E) is supported.

[Shylock]

We use the newest kaldi from github， and can you tell me which version of kaldi you are using?
We refer to the lf-smbr written by vimal's github in “https://github.com/kaldi-asr/kaldi/pull/2140/commits/0bacc8371186561d8c6752cd21e67d7aaee615bb”

[Srikanth M R]

I recently test it with 6e633c9a from April 13, and it is working fine for me. I tested with CUDA 11.3 and Pytorch 1.11.

It is hard for me to find the exact bash script that does lf-smbr training. Could you please share the name of the file that I should be looking at?

[Shylock]

Thanks Srikanth for your kind reply.
I'm very busy these days, and i'm so sorry it took so long to get back to your message
chain-smbr is open on vimal's github, you can get the code from: https://github.com/vimalmanohar/kaldi, and the branch is named "chain-smbr-fixed"

Best wishes,
Shylock

[Shylock]

Dear Srikanth，
I want to know the CER gap between Kaldi(chain) and Pkwrap. Could you show me the comparation between the above two?

Thanks
Shylock

[Srikanth M R]

Hi,

I never checked CER. But the WER comparisons are here in the original report. You can also check the WERs in the recipes provided.

I have never tried lf-smbr, and it might be difficult to integrate it here since it is not there in the original Kaldi repository. Assuming the license for the lf-smbr code is the same as that of Kaldi (I think it is), I can help you demonstrate how to integrate it directly in a branch of this repo.

Thank you,
Srikanth

[Shylock]

Ok, Thanks.
We will try Pkwrap after some time, and we can do more meaningful things with this Pkwrap tool.