Closed zhang-jian closed 6 years ago
Hi,
Parallel WaveNet in my implementation does not apply scaling and shifting in the last flow. So that we need scaling and shifting to get x = z^(n)
.
Not only x
but also means and log scales are needed to calculate loss. That is why I implement as this.
Sorry for the late reply, it is clear now. Thanks.
Hi,
Thanks for sharing.
It seems that there is an extra sampling step after the student network at https://github.com/dhgrs/chainer-ClariNet/blob/master/StudentGaussianIAF/net.py#L58 to get student samples. However,
Algorithm 1
in https://arxiv.org/pdf/1807.07281.pdf uses z from the last IAF flow as student samples. What is the motivation of making such change?Thanks,
Jian