Open fritzo opened 6 years ago
Gosh I haven't looked into the details of implementing .sample()
and .log_prob()
. Is the bessel function needed for sampling or for scoring?
I believe it is required for scoring. Sampling can be done without Bessel functions.
This is a good paper to read from https://www.jstor.org/stable/2346732?seq=1#page_scan_tab_contents.
Will this not require an approximation / implementation of the Bessel function, which I think is not available in PyTorch? However, MKL and CUDA supports these as special functions, so I think a PR with Bessel would be required first.