Visual_Speech_Recognition_for_Multiple_Languages icon indicating copy to clipboard operation
Visual_Speech_Recognition_for_Multiple_Languages copied to clipboard

Published 20 hours ago • verified publisher icon mpc001

Version issues

Open jayden-leo opened this issue 2 years ago • 3 comments

Can you please tell me the life version of pytorch you are using, I have some errors with the 2.0 version. Thank you!

jayden-leo avatar Apr 23 '23 01:04 jayden-leo

It works with pytorch 1.13.1, but the result is very bad........

matiter avatar Apr 28 '23 07:04 matiter

Hi @jaydenjudith, can you please provide more information about the errors you're experiencing? I tested the 2.0.0 version and it worked for me. @matiter, can you please clarify which result you were referring to?

mpc001 avatar May 02 '23 17:05 mpc001

Hi, @matiter @jayden-leo, Dr.Ma @mpc001 is right. The version of pytorch (2.0.0) is the key point. When I use pytorch==1.13, only 6.0% WER for SOTA AVSR, but for version 2.0.0 the results align with the paper well. By the way, make sure to use torchaudio just like the original code to extract audio from mp4, other packages like Vidoeclip may cause different results.

dalision avatar Dec 20 '23 07:12 dalision