residual layer should be one shorter about nv-wavenet HOT 5 CLOSED

Hi @zhf459 . We add the dummy tensor because the nv-wavenet CUDA code expects there to be as many residual weights as there are dilated and skip weights, but it throws away the result of the last residual layer. I didn't want users to have to input a dummy tensor themselves.

from nv-wavenet.

thank you, i see.Another problem is about use_tanh_embed. is it used in reference ,I find it used in in this function.

from nv-wavenet.

I'm not sure I understand the question. In the model you can download on the README we use use_tanh_embed=False. The original DeepVoice implementation used use_tanh_embed=True

from nv-wavenet.

I'm going to request this issue be closed. @zhf459 If you have further questions feel free to contact me or open another issue.

from nv-wavenet.

@RPrenger ok~,thanks!

from nv-wavenet.