Closed fritzo closed 4 years ago
These will benefit from https://github.com/pytorch/pytorch/issues/20343 and https://github.com/pytorch/pytorch/pull/31278
These will benefit from pytorch/pytorch#20343 and pytorch/pytorch#31278
If it ever gets merged :(
Closing in favor of an asymptotically exact method #2410
This issue proposes two distributions that could enable HMC and reparameterized SVI inference for models with count-valued latents, such as epidemiological compartment models (see design doc).
Models such as SIR and SEIR typically use
Binomial
or overdispersedBetaBinomial
distributions for transitions (or approximate these asCensored(Poission)
or overdispersedCensored(NegativeBinomial)
) andBinomial
likelihoods (e.g. see this prototype). In these cases, theS,E,I
variables are integers. This issue proposes to replace the integerstotal_count
with positive real numbers denoting a binary mixture model over the floor and ceiling oftotal_count
. The semantics of the transition and likelihood distributions differ in that transition will be real->real, but likelihood will be real->integer.ContinuousBinomial
likelihoodThis is a simple two-component mixture model. Here's a sketch:
ContinuousBetaBinomial
with reparametrized rsampleThis uses a mixture model for density but also implements
.rsample()
, which is complex. Here's a sketch:Questions
Tasks
ContinuousBinomial
likelihood with.log_prob()
ContinuousBetaBinomial
transition distribution with.log_prob()
and.sample()
. This is sufficient for HMC inference.ContinuousBetaBinomial.rsample()
. This is required for SVI inference.