ritheshkumar95 / energy_based_generative_models Goto Github PK

when I run the MCMC evalulations on CIFAR data:
I got the error: FileNotFoundError: [Errno 2] No such file or directory: '/Tmp/kumarrit/cifar_samples'
And I noticed that : NOTE: This requires cloning the TTUR repo in the current working directory (https://github.com/bioinf-jku/TTUR).
I download the repo. But I didn't find the file.
Can you show me how I can get the directory: '/Tmp/kumarrit/cifar_samples'?
Look forward to you reply ! Thanks!

you use MCMC on random variable z, which means you think the energy function of z is E(G(z)), where E(x) is the energy function of x and G(z) is generator. But why?

We have p_{\theta}(x) = e^{-E(x)}/Z is a energy distribution of x. If we transform it with x=G(z), we get p(z)=e^{-E(G(z))}/Z * abs(det(dG/dz)), where dG/dz is Jacobian of G. The result is not equal e^{-E(G(z))}/Z.

compared with wgan-gp or wgan-div, your new GAN has an additional I(X,Z) term on generator loss. This term may prevent generator from mode collapse.

As we known, wgan-gp or wgan-div can be trained successfully without I(X,Z). But as your derivation in your paper, I(X,Z) is indispensable.

Therefore, how can we understand the success of wgan-gp or wgan-div under your framework ?

Could you post a requirements.txt or an equivalent package to know which version of the various packages are expected?

HI，recently，i want to use MINE to estimate the KL between Gaussian distribution and Laplace distribution。I have noted that how to input when i want to estimate KL（P（XY）||P（X）P（Y））。But I dont know how to input when i want to estimate KL(P(X)||P(Y))，i chosen Gaussian distribution and Laplace distribution，just got N samples from them，do i need to shuffle their sample values？

Good job! But I have some problems:

1. how if we use hinge loss as EnergyModel loss? Gradient penalty is actually good but we know with gradient penalty is slow.

2. Can you show more final demo pictures in github? I really want to see the whole results on cifar10 and celeba.

thanks for such a good job.

Hi, I find the MALA sampling is commented off here: https://github.com/ritheshkumar95/energy_based_generative_models/blob/master/scripts/train/functions.py#L10.
And, I find it only shows up in the testing. So, do you mean it is not helpful during the training?

As we know I(X,Z) = KL(p(x,z)||p(x)p(z)). So why do you estimate mutual information by JSD rather than KL maximization?

f-GAN also gives us
KL(p(x)||q(x)) = max_T E_{x~p(x)} [T(x)] - E_{x\sim q(x)} [e^{T(x)-1}]
I think it is a more natural and more reasonable choice than JSD ?