search

keithito / tacotron

A TensorFlow implementation of Google's Tacotron speech synthesis with pre-trained model (unofficial)
MIT License
2.96k stars 956 forks source link

what is the expected loss after 250k on LJ Speech Dataset? #99

Open alanbekker opened 6 years ago

alanbekker commented 6 years ago

after 140k I'm stuck around 0.08

yyt233 commented 6 years ago

You and I are in the same condition.

keithito commented 6 years ago

That's expected. By around 500K steps, it should drop slightly more, to about 0.077. It might continues to decrease after that point, but the audio doesn't get noticeably better.

alanbekker commented 6 years ago

could you please specify which hyper parameters did you use during optimization?

On Fri, Jan 5, 2018 at 7:47 AM, Keith Ito notifications@github.com wrote:

That's expected. By around 500K steps, it should drop slightly more, to about 0.077. It might continues to decrease after that point, but the audio doesn't get noticeably better.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/keithito/tacotron/issues/99#issuecomment-355478054, or mute the thread https://github.com/notifications/unsubscribe-auth/AE_DSUsu6oUCZnYJcjbjE-jEuA1SFqudks5tHbeHgaJpZM4RRY51 .

keithito commented 6 years ago

I used the defaults.

SoulFireMage commented 6 years ago

How long does it take to reach 250k steps? I'm on 758 and it's been going 2 hours, at this rate I'll be here for several weeks at least. Is this normal?

keithito commented 6 years ago

No, that's not normal if you're training on a GPU. Please make sure you've installed Tensorflow with GPU support. See: https://www.tensorflow.org/install/