Open austinv11 opened 10 months ago
torch.multinomial
provides a sampler. a full-fledged distribution also requires a log_prob
. however afaik evaluating the hypergeometric function involves nasty special functions that aren't available in pytorch
possibly relevant links: https://github.com/google-deepmind/torch-cephes https://github.com/pytorch/pytorch/issues/76324 https://github.com/pytorch/pytorch/issues/3877 https://github.com/pytorch/pytorch/issues/36036
Ah, that's unfortunate. Thanks for the info! Feel free to close.
Issue Description
In my work I need to develop a generative model that makes use of the Hypergeometric distribution. I believe
torch.multinomial
without replacement is equivalent? But it is of course, incompatible withpyro.sample
statements.Suggestions for alternative distributions would also be appreciated.
Thanks for your consideration!