the logPx log_var is zero

[benlaitang]

i am curious anout https://github.com/JeremyCCHsu/vae-npvc/blob/master/model/vae.py#L124
logPx = tf.reduce_mean(
GaussianLogDensity(
slim.flatten(x),
slim.flatten(xh),
tf.zeros_like(slim.flatten(xh))),
if log_var is constant 0, leading to loss GaussianLogDensity is equivalent to MSE？

[JeremyCCHsu]

Yes, it's intended. Constant variance prevents overfitting.

[benlaitang]

Yes, it's intended. Constant variance prevents overfitting.

I noticed that D_KL loss https://github.com/JeremyCCHsu/vae-npvc/blob/master/model/vae.py#L113
and logPx https://github.com/JeremyCCHsu/vae-npvc/blob/master/model/vae.py#L120 is related ~N(0,1) where u is 0 and log var is 1, this is not same to paper. it is just a trick for training?
Actually, I finished the VAE code for training multiple speaker, but it did not work, I use the KL loss+ MSE loss.

[JeremyCCHsu]

I don't quite understand your question... Var=0 --> log var = 0. Actually, tuning log var could help in some cases.

update: typo: var=1 --> log var =1

[benlaitang]

I don't quite understand your question... Var=0 --> log var = 0. Actually, tuning log var could help in some cases.

log var =0 is var = 1.
Can you help check the VAE loss, give me some advice? Now it did not work.