search

yuhogun0908 / MISOnet

Unofficial Multi-microphone complex spectral mapping for utterance-wise and continuous speech separation(MISO-BF-MISO)
MIT License
51 stars 9 forks source link

training environment issue #3

Closed ermu-tech closed 2 years ago

ermu-tech commented 2 years ago

Hi! Your code is very clear, but I think I have encountered some issues about the version of packages while running your code. Could you please share your training environment of your code and the versions of each package? Thank you!

ermu-tech commented 2 years ago

By the way, could you please share the dataset you use to run the code? Thank you !

jzi040941 commented 2 years ago

Hi, I've added requirements.txt file and updated README for environment setting Thanks!

yuhogun0908 commented 2 years ago

Hi You can refer to the dataset github of the paper.

ermu-tech commented 2 years ago

Thank you for your reply! It's very useful. But there was another problem: in run.py line 18, 'from tester import Tester' doesn't work, maybe you missed a 'test' file in your code?

ermu-tech commented 2 years ago

Hi! I took a look at your code, and now I have few other questions to consult：

  1. From the way it's named, I think the file 'MISOnet/dataloader/REVERB_2MIX.py' is used to add reverb to mix signal, and file 'MISOnet/dataloader/RIR_mixing.py' is used to add RIR to mix signal. But after looking though the code, I'am not sure what these two files are used for, could you please give some explanation in English? (It's a pity that I can't understand korean notes very well)
  2. It seems that your code just MISO1(separation network), do you have the whole MISO1-BF-MISO3 system code and would like to share it? Please excuse me if I misunderstood your code. Looking forward to your reply! Thank you!
jzi040941 commented 2 years ago

Tester moudule is exactly same meaning as it's name, it's needed for testing You can just comment out that line

Answer for 1: At the time that we begin developing this repo we didn't have training dataset such as sms-wsj, libri-css So we suppose to generate temporary dataset before we generate sms wsj dataset, that's why we implemented rir mixing If you have sms-wsj dataset we dont need those thing

Answer for 2: We are still developing system but not fully implemented whole system, We will make commit When the code is stabilized Stay tune thanks

By the way, if you able to implement some part of systen feel free to make fork repo and make full request thanks

ermu-tech commented 2 years ago

Thank you for your reply! Can I assume that the two file 'REVERB_2MIX.py' and ' RIR_mixing.py' are used to generate temporary dataset? Actually I don't have sms-wsj, libri-css dataset, either. So do you mind to share which dataset you used to generate temporary dataset? Maybe it's much easier to run your code with that. Thank you!

ermu-tech commented 2 years ago

Hi! I'm not sure if you have seen my message above. I still wonder if you could share your dataset. Sorry to interrupt you !

yuhogun0908 commented 2 years ago

Sorry to late answer.

Answer : Can I assume that the two file 'REVERB_2MIX.py' and ' RIR_mixing.py' are used to generate temporary dataset?

--> I used REVERB_2MIX.py, RIR_mixing.py dataset temporarily to implement and verify code. To be more specific, RIR mixing was performed through simulation through Matlab. RIR mixing method : Image method Paper : [J. B Allen and D. A. Berkley, "Image method for efficiently simulating small-room acoustics,"J.Acoust. Soc. Am. 65, 943-950(1979)]

Answer : Actually I don't have sms-wsj, libri-css dataset, either So do you mind to share which dataset you used to generate temporary dataset? Maybe it's much easier to run your code with that. --> You can proceed through https://github.com/fgnt/sms_wsj to make sms-wsj and https://github.com/chenzhuo1011/libri_css to make libri-css.

sorry about that temporary dataset was removed after code implementation.