alan

Software in development!! Not yet for external use!!

Installation

Start by removing previous alan or alan_simplified

python -m pip uninstall alan
python -m pip uninstall alan_simplified

And then remove the corresponding repo folders before cloning the new repo.

To install, clone repo, navigate to repo root directory, and use,

pip install -e .

Tests

To run tests, navigate to tests/ and use pytest.

Docs

Read the Docs

Overall example:

See examples/example.py

Meeting TODOs:

• Docs: Friendly overview material (Pyro is a good example).
  • clarify ExtendedImportanceSample.predictive_ll.
    • How does it deal with the N samples?
    • How does it deal with the LL for the training data?
• Think carefully about extended Plate errors.
  • e.g. if you try to extend a prior with plated parameters.
• Timeseries:
  • tests (Kalman filter)
  • importance_sample
  • extend
• MovieLens/bus experiments with new code, with VI / RWS / QEM==Natural RWS (see example/example.py):
  • small scale (usual subsampling)
  • large scale using computation_strategy=Split(...)
• User-facing Marginals.ess

Long-run TODOs:

• Friendly error messages:
  • Marginals/moments make sense for variables on different plates if they're in the same heirarchy.
• Enumeration:
  • Enumeration is a class in Q (like Data), not P.
• A Samples class that aggregates over multiple Sample in a memory efficient way.
  • Acts like it contains a list of e.g. 10 Samples, but doesn't actually.
  • Instead, it generates the Samples as necessary by using frozen random seed.

Ideas:

• Remove checkpointing from reduce_Ks?
• Numerically stable logsumexp_dims, by adding t.finfo(lp.dtype).eps before log
• Check P: Timeseries, Q:Plate/Timeseries, but not P:Plate, Q:Timeseries, or P:Plate, Q:Data.
• Check timeseries permutation is the right size.
• Simplify sampler.resample_scope, by noting that:
  • all variables in scope are either parameters (no K's)
  • random variables drawn from a dist (unique K)
  • random variable drawn from a group (non-unique K).
  • can reason about these more easily if we e.g. pass in varname2groupvarname.
• importance_sample.dump should output tensors with the N dimension first?
• latent moments for linear_gaussian_latents
• tests for mixture distributions.
• remove the random init from most of the tests.
• tests for extended_sample
  • TestProblem takes extended_platesizes and predicted_extended_moments as arguments.
  • some extended moments are exactly equal to importance sampled moments (i.e. unplated moments + first part of plated moments).
  • other extended moments aren't exactly equal.
  • instead, the extended moments are a function of importance_sample moments.
  • TestProblem takes
  • predicted_extended_moments is a function that takes an importance sample, and returns mean + variance of moment.
  • should really compare moments to
• QEM distributions:
  • Categorical
  • Testing