Skip to content

Releases: borodark/exmc

I

07 Feb 07:41
@borodark borodark
v0.1.0
37bb04e

Choose a tag to compare

View all tags
I Latest
Latest 
v0.1.0

Take I
Assets 2
Loading

Four Chains, Zero Locks

07 Feb 07:35
@borodark borodark
α
37bb04e

Choose a tag to compare

View all tags
Four Chains, Zero Locks

Exmc and ExmcViz: a probabilistic programming framework and native MCMC visualization toolkit for the BEAM.

Exmc — Probabilistic Programming for Elixir

Model Specification

  • Declarative IR with Builder API: rv, obs, det nodes
  • 9 distributions: Normal, HalfNormal, Exponential, Gamma, Beta, Uniform, StudentT, Cauchy, LogNormal
  • Hierarchical models via string parameter references (sigma: "sigma")
  • Vectorized observations — pass a tensor to Builder.obs, reduction is automatic
  • Rewrite pipeline with composable named passes

Sampling

  • Full NUTS sampler (Betancourt 2017) with multinomial tree building
  • Stan-style three-phase warmup: step-size search, initial adaptation, final mass matrix
  • Diagonal mass matrix estimation with windowed adaptation
  • Dual averaging step-size tuning
  • Automatic constraint transforms with Jacobian correction (log, softplus, logit)
  • Automatic non-centered parameterization for hierarchical Normals
  • Parallel multi-chain sampling — sample_chains/3 compiles once, dispatches via Task.async_stream
  • Live streaming: sample_stream/4 sends samples to any process as they're drawn

Diagnostics & Model Comparison

  • Effective sample size (ESS) via initial positive sequence estimator
  • Split R-hat convergence diagnostic
  • WAIC and LOO-CV model comparison
  • Prior and posterior predictive sampling

Backend

  • Nx tensors with automatic EXLA JIT detection — value_and_grad is compiled when EXLA is available, falls back to BinaryBackend transparently
  • Erlang :rand for PRNG (1000x faster than Nx.Random on BinaryBackend)

Test Suite: 123 tests (11 doctests + 112 tests), 25 integration tests

ExmcViz — Native MCMC Diagnostics

ArviZ-style visualization through Scenic's OpenGL pipeline on a true-black OLED palette.

Visualization Types

  • Dashboard (show/3) — trace plot, histogram, ACF, summary panel per variable; energy plot at bottom; dynamic proportional layout
  • Pair plot (pair_plot/2) — k x k grid: histogram diagonal, scatter lower triangle, Pearson correlation upper triangle
  • Forest plot (forest_plot/2) — horizontal HDI intervals (50% + 94%) with posterior mean dots
  • Live streaming (stream/3) — real-time dashboard updated via GenServer coordinator buffering 10 samples per flush

Features

  • Multi-chain overlay with 10-color warm spectrum
  • Divergence markers (red circles on trace plots)
  • Energy diagnostics (overlaid marginal + transition histograms)
  • Dynamic viewport-proportional layout — priority: trace 50% > histogram+ACF 30% > summary 20% > energy 8%
  • 4K portrait optimized (2160 x 3840)

Test Suite: 34 tests

Architecture

35 architectural decisions documented in DECISIONS.md, from :rand vs Nx.Random (D16) to auto-NCP (D32) to compile-once parallel chains (D35).

License

Loading