Why is the Loss (L^x + L^z) so small?

[lovecambi]

It is about 70, lower than most reported results.

I also find one bug. I think you are misleading by the Eq. (12)., the equation is used to compute each element of vector z.

kl_terms[t]=0.5*tf.reduce_sum(mu2+sigma2-2*logsigma,1)-T*.5 # each kl term is (1xminibatch)

should be

kl_terms[t]=0.5*tf.reduce_sum(mu2+sigma2-2*logsigma-1,1) # each kl term is (1xminibatch)
# kl_terms[t]=0.5*tf.reduce_sum(mu2+sigma2-2*logsigma,1)-T*z_size*.5 # alternatively

Or the kl term will blow up with large z_size.

Another issue is the mnist data in your code is not binarized. But it won't make much difference.

[lovecambi]

I know the reason. The sampled gaussian noise should be put into the function sampleQ.

[ericjang]

Thanks for the catch, sorry I didn't address this earlier. Would you mind submitting a pull request to fix this?

…

On Mon, Feb 27, 2017 at 7:47 AM, lovecambi ***@***.***> wrote: I know the reason. The sampled gaussian noise should be put into the function sampleQ. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#2 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAacMYqmE4L-gJbz2CHn6EHLVSd0YEhgks5rgvAjgaJpZM4KcmC1> .