Regarding MNIST-RBM

[gnobitab]

Hi authors!

Thank you for providing the source codes for the excellent work!

I have a question regarding the MNIST-RBM experiments (rbm_sample.py). It seems that all the methods are initialized with a gibbs sampling result (in other words, model.init_dist is changed by first running gibbs sampling for 5,000 steps). Did you try to sample without this initialization? Moreover, since all the methods are initialized by gibbs ( the yellow curve), why do they have a much higher log-MMD than the yellow curve? Finally, could I know what the algorithmic difference is between dim-gibbs (the blue curve) and gibbs (the yellow curve)?

Looking forward to your replay. Thanks in advance!!!