Using flows for density estimation

[pratham16cse]

Thanks for providing time-series compatible normalizing flows in flows file. I am interested in using normalizing flows for density estimation.

I have a couple of doubts about how masks are created in MADE block of MAF: Suppose the input is of shape B x T x D where B = batch size, T = Sequence length and D = Dimension of the sequence and I want to estimate the density at each position (Eq. 2 of MAF Paper). In this case,

1. What should be the input_size to MAF network?
2. Can you give a toy example of what masks will be generated in the MADE block?
3. Also, how the cond_label_size would fit in the whole equation?

Perhaps specifying the shapes of the input and output of each block in flows file would be helpful. Please let me know if more elaboration on the setup is needed.

Thanks,

• Prathamesh

[kashif]

@pratham16cse i would not use these specific modules for general density estimation as they are specific to the conditional estimation of a sequence of multivariate vectors...

I do not remember exactly the variables names right now... I would need to look at the code again...

I believe the flows because they are invertiable do not change the size of the input so all that remains the same... so i belive the input_size might be multivariate dim size and cond_lable size might be the size of the condition vector... I have to check

[pratham16cse]

Thank you for the response @kashif. I am trying to use the flow modules for density estimation in reference to an existing paper. I believe the authors of this paper have used your code from flow modules (link).

I think I have got enough clarification on the question I asked. However, in case you strongly suggest not using modules for density estimation, I would like to take this discussion offline if you are interested. Thanks a lot.

[kashif]

yes lets have a video chat soon and discuss things! I am interested in your work and wish to also understand it further. Kindly send me an invite at [email protected] for some reasonable time... I would love to chat!