I notice that you use GatedDense in self.q_z_layers to encode q(z | x):

Is there any reason to use GatedDense rather than a simple MLP? Thanks!

Hi, in line 78~80 of VAE.py, when calculating the KL divergence of VampPrior p(z) and posterior q(z|x), we put z_q into N(z_p_mean, z_p_logvar) and N(z_q_mean, z_q_logvar) then calculate the difference of log outputs of two Gaussian function.

Since we have already get mean and variance of prior and posterior, can we directly calculate KL divergence of two Gaussian distribution? i.e.

This is because in line 226, z_q is just drawn from N(z_q_mean, z_q_logvar):

Trying to get this working today, but I'm getting the following error:

python experiment.py --dataset_name=cifar10

load data
Files already downloaded and verified
create model
Traceback (most recent call last):
File "experiment.py", line 160, in
run(args, kwargs)
File "experiment.py", line 124, in run
from models.VAE import VAE
File "/workspace/vae_vampprior/models/VAE.py", line 20, in
from Model import Model
ImportError: No module named 'Model'

Any ideas?

In the generative part of the forward pass when computing the joint density p(x | z1, z2):

I suppose there is a self.reparameterize(z1_p_mean, z1_p_logvar) missing, i.e.:

# p(z1 | z2)
z1_p_mean, z1_p_logvar = self.p_z1(z2_q)
z1_p = self.reparameterize(z1_p_mean, z1_p_logvar)

# x_mean = p(x|z1,z2)
x_mean, x_logvar = self.p_x(z1_p, z2_q)

Let me know if I am wrong :-)
Very good code anyways! 👍

I'm unsure of the role the pseudo inputs play in generation using a Vamp Propr. This is likey my lack of understanding of your paper because I find it hard to reconcile these implementation details with your paper (my lack of expertise in this area).

It seems that when you generate, you generate using a subset of the pseudo inputs as per:

means = self.means(self.idle_input)[0:N] z_sample_gen_mean, z_sample_gen_logvar = self.q_z(means) z_sample_rand = self.reparameterize(z_sample_gen_mean, z_sample_gen_logvar)

I understand how to generate from these pseudo inputs using reparameterization but how would I sample from the aggregated prior itself? After all, it is this prior that one would normally sample from in a VAE after all (though it wouldn't be aggregated in the vanilla VAE).