bgmCompare: persistent sweep state and current-state normalizer cache#167
Merged
Conversation
…weights, and residuals Add CompareSweepState, owned by run_gibbs_sampler_bgmCompare and threaded by reference through the Gibbs update step. It holds the double observation matrices (whole and per group), the group-specific effective pairwise weights, and the residual matrices obs_double[g] * pairwise_group[g]. The state replaces per-use rebuilds: - the AM pairwise sweep and the proposal-sd tuner no longer rebuild the weight and residual matrices each sweep; accepted moves maintain them; - log_pseudoposterior_main_component reads rest scores from the residual matrices instead of converting observations and running one GEMV per group per evaluation; - the indicator-flip ratios read current rest scores from the residual matrices and adjust them by the change in the effective weight, removing the per-call conversions and GEMVs; accepted pairwise flips update the state in place; - the NUTS path receives the pre-converted observation matrix and triggers one state rebuild per iteration; - imputation refreshes the state once per pass. Verification: test-bgmCompare.R passes (40, 42 with NOT_CRAN). Same-seed single-chain regression against the previous commit: NUTS-with-selection and AM-with-imputation fits are bitwise identical; AM fits without imputation differ only by residual-maintenance float reordering (max sample diff 8e-8, posterior-mean diff 3e-8); indicator draws bitwise identical in all cells.
Add a per-variable, per-group cache of the pseudolikelihood log-normalizer sums to CompareSweepState, and metropolis_step_cached(), a metropolis_step() variant that takes the current-state log-posterior as a value and evaluates only the proposal (same random-number stream). The main-effect and pairwise-effect component functions accept optional cached normalizer sums for the current state and optionally return freshly computed ones. Two shared helpers, metropolis_update_main_effect_cached and metropolis_update_pairwise_effect_cached, reconstruct the current-state value from the cache (computing it on a miss), evaluate the proposal in full, and on acceptance store the proposed-state normalizer sums; the Metropolis sweeps and the proposal-sd tuner call the same helpers. This removes the second full evaluation per proposal. Invalidation: accepted main and pairwise moves store the proposed-state sums; accepted indicator flips invalidate the affected variables; weight rebuilds (NUTS updates, graph initialisation, imputation refresh) invalidate every variable. Verification: test-bgmCompare.R passes (42 with NOT_CRAN). Same-seed single-chain regression against the previous commit: all four fits (AM without selection, AM with selection, AM with selection and imputation, NUTS with selection) are bitwise identical, including indicator draws; AM fits run 25-45% faster.
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## develop #167 +/- ##
===========================================
+ Coverage 90.98% 91.07% +0.08%
===========================================
Files 104 104
Lines 14913 14950 +37
===========================================
+ Hits 13569 13615 +46
+ Misses 1344 1335 -9 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Two commits removing the main redundancies in the bgmCompare sampler. Follow-up to #166.
Persistent per-group sweep state
CompareSweepState, owned by the Gibbs runner and threaded by reference, holds the double observation matrices (whole and per group), the group-specific effective pairwise weights, and the residual matricesobs_double[g] * pairwise_group[g]. It replaces per-use rebuilds: the AM pairwise sweep and proposal-sd tuner no longer reassemble weight/residual matrices each sweep (accepted moves maintain them in place), the main-effect pseudoposterior reads rest scores from the residuals instead of converting observations and running one GEMV per group per evaluation, and indicator-flip ratios adjust cached rest scores by the effective-weight change. The NUTS path triggers one state rebuild per iteration; imputation refreshes the state once per pass.Current-state normalizer cache
The sweep state additionally caches the per-variable, per-group pseudolikelihood log-normalizer sums, and
metropolis_step_cached()takes the current-state log-posterior as a value (same random-number stream), so each Metropolis proposal evaluates only the proposal instead of both states. Accepted moves store the proposed-state sums; indicator flips invalidate the affected variables; weight rebuilds (NUTS updates, graph initialisation, imputation refresh) invalidate every variable.Verification
Benchmarks (installed -O2 builds vs the base branch; n = 800, V = 16 ordinal, 3 groups, single chain, min of 2)
The AM gains grow with n and the number of groups (the removed work — observation conversions, per-group GEMVs, current-state re-evaluations — scales with both). NUTS benefits only through the Metropolis indicator flips.