search

JaebeomKang / HTD_BSRNN_vocal

Graduation project at Yonsei Univ.
6 stars 0 forks source link

Results about BSRNN #1

Open ycyang133 opened 8 months ago

ycyang133 commented 8 months ago

Hi @JaebeomKang,

Thanks for your great work! I am confused about the SDR results. Does the orange indicates the original BSRNN? Why did the SDR decrease after about 30 epochs? Thank you so much for your attention.

JaebeomKang commented 8 months ago

Hi, Yichen!

Thanks for your interest in my project.

Well, honestly I also surprised when I got the results of BSRNN. So I tried multiple times with same conditions, but I got the same results.

I guess that the BSRNN used the single FC layer for encoding-decoding and it may cause the overfitting problems easily. (Decreased performance after several epochs often implies overfitting)

But I can assure you that I didn't make any changes to the open source code.

Sorry for the unclear answer:(

If you have any other questions, feel free to ask me.

Have a nice day!

2024년 1월 29일 (월) 오후 3:18, Yichen @.***>님이 작성:

Hi @JaebeomKang https://github.com/JaebeomKang,

Thanks for your great work! I am confused about the SDR results. Does the orange indicates the original BSRNN? Why did the SDR decrease after about 30 epochs? Thank you so much for your attention.

— Reply to this email directly, view it on GitHub https://github.com/JaebeomKang/HTD_BSRNN_vocal/issues/1, or unsubscribe https://github.com/notifications/unsubscribe-auth/BCCJJ7FW24KPXDXXJIFWXALYQ45MVAVCNFSM6AAAAABCO3OLQGVHI2DSMVQWIX3LMV43ASLTON2WKOZSGEYDINZTGYZTOMA . You are receiving this because you were mentioned.Message ID: @.***>

ycyang133 commented 8 months ago

Thank you for your prompt answer! Yes, it seems overfitting occurs if these curves are obtained on the validation or test set. Maybe a deeper network or a better data augmentation way is needed to relieve this. Thank you so much again. By the way, has this work been published?

JaebeomKang commented 7 months ago

Not yet. Can I ask you Why?

2024년 2월 1일 (목) 오후 2:09, Yichen @.***>님이 작성:

Thank you for your prompt answer! Yes, it seems overfitting occurs if these curves are obtained on the validation or test set. Maybe a deeper network or a better data augmentation way is needed to relieve this. Thank you so much again. By the way, has this work been published?

— Reply to this email directly, view it on GitHub https://github.com/JaebeomKang/HTD_BSRNN_vocal/issues/1#issuecomment-1920526191, or unsubscribe https://github.com/notifications/unsubscribe-auth/BCCJJ7BG54P2F34IJEEBFKTYRMPRFAVCNFSM6AAAAABCO3OLQGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTSMRQGUZDMMJZGE . You are receiving this because you were mentioned.Message ID: @.***>

ycyang133 commented 7 months ago

Just because it seems already in good shape and ready for publication.