search

jliu / graph-normalizing-flows

Code for Graph Normalizing Flows.
59 stars 9 forks source link

Flow with H_T as Gaussian Rather Than H_0? #1

Open RuihongQiu opened 4 years ago

RuihongQiu commented 4 years ago

Hi Jenny,

I find it really interesting of this GNF work. And I have a question in the model design for the graph generation in Figure 2 of the original NeurIPS 2019 paper. image

I think the procedure of flows is different to the normalizing flows paper in ICML 2015. According to this paper, the H_0 should be close to the Gaussian prior rather than H_T as following: image

Could you give a simple explanation?

Thank you so much!

nkuweili commented 3 years ago

I think that setting H0 closer to the Gaussian prior rather than HT(as previous) may be convenient for sampling during generation process. Sampling samples from H0 and then going through reversible GNF and Decoder will generate a set of new graphs (like the dotted line in Figure 2 of the paper). This is my understanding