View source
deletemove
  • Contributors
    1. Loading...
Author Commit Message Date Builds
250+ commits behind master.
Mike Hughes
FIX deletemove final fixes... made empty-topic delete possible even when merge occurs, and removed errant embed statement.
Mike Hughes
ENH deletemove now works alongside mergemove. HACK: when a merge is accepted, no deletes can occur
Mike Hughes
ENH deletemove improvements: works with Gauss/Mult and HDP/DP, and has specific last-minute ability to delete empty topics... TODO integrate with mergemove
Mike Hughes
FIX DeletePlanner wont print everything about first/second tier eligible ids, just the top few
Mike Hughes
FIX MergePlanner now always ignores correlations below 0.01 as signal to merge... Should be a reasonable thing since it prevents bad attempts using a very small correlation
Mike Hughes
ENH deletemove seems to be working, after huge fix that better tracks sumLogPiRem. Now tracking failures at each eligible comp uid, and removing comps from consideration after a few strikes. TODO: revisit construction... success rates still not where we'd like them.
Mike Hughes
FIX selectSubsetLP now correctly takes a subset of the ElogPi vector, *not* all of it
Mike Hughes
ENH improved deletemove on target set. Seems to actually work on toy data, but needs some thought about avoiding trying the same comp too many times, how to select eligible candidates, etc.
Mike Hughes
ENH HModel calc_evidence can take just Data obj (and optional LPkwargs), and will compute the ELBO directly
Mike Hughes
Merge branch 'master' into deletemove
Mike Hughes
ENH temporary birth-merge alg with print statements
Mike Hughes
ENH HDPDir allocmodel now has selectSubsetLP method, for making small-sized target data given a much bigger set of local params
Mike Hughes
FIX WordsData bug fixed, where incrementally aggregated data had erroneously retained cached doc-token count matrix
Mike Hughes
ENH deletemove module now has implemented functionality for target-based delete moves. Moves are planned, appropriate data collected, and then decisions made based on the target set.
Mike Hughes
ENH delete move progress, now have solid routine for checking efficacy on example data
Mike Hughes
ADD delete testcase
Mike Hughes
ENH DeleteMove for whole datasets (where we can afford touching all resps) is ready. Exploring several options for resp construction now.
Mike Hughes
ENH added deletemove method for *whole* dataset analysis.