FR Discrete compartmental models for epidemiology

fritzo commented 4 years ago

This issue tracks tasks related to discrete compartmental models. See design doc for motivation.

Modeling

[x] #2410 Initial prototype using HMC + enumeration
[x] #2437 Implement a framework in pyro.contrib.epidemiology
[x] #2446 Implement an SEIR model @fritzo
[x] #2450 #2451 Generalize to overdispersed noise via BetaBinomial @fritzo
[x] #2457 Handle observations accumulated over an interval
[x] #2460 Allow stochastic init state and start date estimation
[x] #2466 Add a regional model
[x] #2474 Add a CoalescentTimes distribution
[x] #2514 Eliminate need for ComparmentalModel.transition_bwd()
[x] #2517 Add a model with time-varying global parameters
[x] #2558 ~~#2468~~ Incorporate phylogeny observations into the likelihood (see OverdispersedSEIRModel)
[x] #2558, #2559 Add example of branching flows #2552
[ ] Add example of causal intervention (ideally interactive on nbviewer)
[ ] Add example of cyclic flows #2552
[ ] Support enumeration of discrete globals, e.g. to condition on multiple phylogenetic samples

Inference

[x] #2428 Use block diagonal mass matrix after #2425 is merged
[x] #2429 Fix global parameters that are easy to physically measure
[x] #2430 Try a frequency-smoothed DCT transform @fritzo
[x] #2434 Try larger enumeration windows @martinjankowiak
[x] #2452 Use SMC to generically initialize to a feasible trajectory
[x] #2465 Try structured adaptation matrices in HMC @fehiepsi
[x] #2492 #2493 Try a wavelet transform @martinjankowiak
[x] #2495 Implement a SplitReparam to split off low-frequency components of Haar reparams
[x] #2510 #2513 #2522 Try continuous relaxed inference
[x] #2498 Make all Binomial distributions overdispersed
[x] #2523 Support Haar reparameterization for non-compartmental time series
[x] #2529 Try SVI on the continuous relaxed model
[ ] Try IWAE with DReG gradients, compare with ELBO (see #2598)
[ ] Try one-step-lookahead SMC
[ ] Try a windowed modified DCT transform

Engineering

[x] #2431 Fix prediction bug when using --dct
[x] #2433 Fix cubic spline weights @martinjankowiak
[x] #2493 Use .with_cache(1) in DiscreteCosineTransform (see https://github.com/pytorch/pytorch/pull/36882 )
[ ] Make SafeLog jit-script-compatible (see comment in hmm.py)

Tidying up for first release DONE

[x] #2499 combine sir.py + seir.py -> models.py
[x] #2553 Add a notebook tutorial introducing the module purpose and usage
[x] #2558 ~~#2518~~ Add a notebook tutorial on regional models
[x] #2563 https://github.com/pyro-ppl/sandbox/pull/13 Set good default values for inference

martinjankowiak commented 4 years ago

great task list. i'm going to try to put together something so we can try larger enumeration windows

fritzo commented 4 years ago

Re: splines, I think we will need different spline weights for differently sized windows. Cubic splines require at least four points. Each point added beyond four gives one new degree of freedom on the coefficients to be solved (the other three degrees of freedom are used by the boundary conditions). One natural way to constrain these extra degrees of freedom in spline coefficients is to minimize the sum of squares of cubic coefficients, i.e. a simple notion of maximally smooth.

fritzo commented 3 years ago

Closing in favor of finer-grained issues.

pyro-ppl / pyro