search

r9y9 / wavenet_vocoder

WaveNet vocoder
https://r9y9.github.io/wavenet_vocoder/
Other
2.33k stars 500 forks source link

are these the same channel numbers in original paper? #82

Closed hrbigelow closed 5 years ago

hrbigelow commented 6 years ago

Hi,

Thanks so much for such a thorough implementation of WaveNet. I just wondered, do your default hparams channel sizes reflect those used in the final version of the original paper? I'm new to the field of AI (coming from Biology research) and am puzzled why the authors wouldn't provide a short paragraph listing all of their final hyperparameters, so others could attempt to reproduce their results. But, maybe I'm missing something. Is their implementation available somewhere?

Thanks very much!

azraelkuan commented 6 years ago

u can check ibab's wavenet, that is origin.

r9y9 commented 6 years ago

As far as I remember correctly, they don't provide complete implementation details, Taco2 and Parallel WaveNet papers provide some details more than the original paper though. I'm not trying to replicate their papers exactly, trying same ideas instead.

hrbigelow commented 6 years ago

Hi Ryuichi,

Thanks for checking into it. I did read ibab/tensorflow-wavenet, and he mentions in the readme that they don't specify channel numbers in the original paper. Tacotron2 does mention they use 3 x 10 layer blocks, but don't seem to mention channel numbers either.

In your experience, when you find a paper that doesn't give enough details, do you have some technique for figuring out how to replicate their results? It seems it would be hard to do research otherwise...

Thanks again,

Henry

On Sun, Jun 17, 2018 at 12:03 AM, Ryuichi Yamamoto <notifications@github.com

wrote:

As far as I remember correctly, they don't provide complete implementation details, Taco2 and Parallel WaveNet papers provide some details more than the original paper though. I'm not trying to replicate their papers exactly, trying same ideas instead.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/r9y9/wavenet_vocoder/issues/82#issuecomment-397859225, or mute the thread https://github.com/notifications/unsubscribe-auth/AFG3FuM9bNyKZeY-ZoAtPYW8r7xjS0ZWks5t9f89gaJpZM4UqiyZ .

r9y9 commented 6 years ago

Trying to find out missing details is fun for me. I'm just trying a lot.

hrbigelow commented 6 years ago

Cool! Yeah, I understand. By the way, are you the professor at Waseda? I have a friend at Ascent robotics doing machine learning and also I lived in Tokyo last year for three months.

Cheers,

Henry

On Thu, Jun 21, 2018 at 6:50 AM Ryuichi Yamamoto notifications@github.com wrote:

Trying to find out missing details is fun for me. I'm just trying a lot.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/r9y9/wavenet_vocoder/issues/82#issuecomment-399110218, or mute the thread https://github.com/notifications/unsubscribe-auth/AFG3FmJgI0YzFey2rgZT12hWfgKhg6hyks5t-6SngaJpZM4UqiyZ .

r9y9 commented 6 years ago

No. See https://www.linkedin.com/in/zryuichi/?locale=en_US about me.

stale[bot] commented 5 years ago

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.