Open Jamesswiz opened 5 years ago
I can see the implementation computes KL div between q(z|x,c) and p(z)=N(0,I) but the original formulation has conditional prior p(z|c).
Can you explain why you are still using zero mean unit Gaussian prior?
I have the same question. And in your blog of vanilla VAE, you have this code comment # D_KL(Q(z|X) || P(z|X)); calculate in closed form as both dist. are Gaussian . The P(z|X) should be P(z) ?
# D_KL(Q(z|X) || P(z|X)); calculate in closed form as both dist. are Gaussian
I can see the implementation computes KL div between q(z|x,c) and p(z)=N(0,I) but the original formulation has conditional prior p(z|c).
Can you explain why you are still using zero mean unit Gaussian prior?