New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Melspectrogram cant be set 'trainable_fb=False' #67
Comments
It's a problem with tf.keras the first solution that comes to mind is to declare MelSpectrogram and set layers as not trainable
you can verify with [<tf.Variable 'melgram/real_kernels:0' shape=(512, 1, 1, 257) dtype=float32>, First two are part of Spectrogram Class [Kernel] and Variable:0 is [fb] due author forgot to set a name for it. you should be able to set as non trainable any of them separately |
@zhh6 Could you provide a test code that checks it like we have in
|
Thanks a lot. I have solved the problem.Just like you do that. |
Why not to use only tf.Variable(..,trainable=False) however my question is ...@keunwoochoi why did you decide to set those variables as "weights" trainable or not? Shouldn't they be fixed? |
That's because one might want to set it trainable to optimize the time-frequency representation. |
This would be one of the things I'd like to fix wrt |
I'm trying to update using tf.stft ... however I think it's going to be impossible set kernels as trainable in that scenario... In fact Spectrogram class would be very short... Do you have any idea or recommendation to that update? |
that'd be great, thanks! and you're right - with tf.stft, one can't it trainable. i just made |
Melspectrogram cant be set 'trainable_fb=False',after I set trainable_fb=False,trainable_kernel=False,but seems like it doesnot work.It is still trainable.
The text was updated successfully, but these errors were encountered: