Skip to content

Latest commit

 

History

History
124 lines (85 loc) · 4.63 KB

NEWS.md

File metadata and controls

124 lines (85 loc) · 4.63 KB
Raw

master (unpublished)

general

  • strongly speed-up contingency table computation for heterogeneous=true and max_k=0/1
  • use the more compiler-friendly stack (introduced in Julia v1.9) instead of hcat for large numbers of columns (if available), introduce fast method for sparse columns
  • improve univariate pvalue filtering
  • remove performance bottleneck in three-way adjust_df
  • catch incorrect usage of 'meta_data_path' as keyword argument

v0.19.2 (latest)

bug fixes

  • force si_HITON_PC to remove non-significant univariate candidates earlier
  • fix edge case where direct usage of the unexposed sparse processing engine could lead to infinite loops
  • fix SparseArrays incompatibility (surfaced by julia v1.9)

v0.19.1

general

  • expose experimental keyword arguments within learn_network for easier testing (subject to change, syntax/functionality should not be relied upon)

bug fixes

v0.19.0

general

  • add support for shared-memory OTU tables (sparse and dense), invoked via the share_data flag in learn_network
  • improve speed and edge case behaviour for non-recursive partial correlations
  • improve freeing of memory (in particular on parallel workers) between major network inference steps
  • improve error reporting on parallel workers

bug fixes

  • fix numerical edge case for recursive partial correlations
  • fix handling of variables with no absences when heterogeneous=true

v0.18.1

bug fixes

  • add hotfix for irregular breaking changes in SimpleWeightedGraphs.jl
  • fix bug that prevented adaptive_clr from correctly propagating samples that were dropped due to numerical edge cases (thanks @jonasjonker for reporting)

v0.18.0

general

  • fast_elim flag now also exposed via learn_network
  • fixed documentation for save_network/load_network to drop outdated .tsv/.csv mention (thanks @ARW-UBT for reporting)
  • update dependency versions

bug fixes

  • fix bug that prevented combining sparse .biom data with non-numeric meta data (issue #20, thanks @ARW-UBT for reporting)

v0.17.0

general

  • Julia versions < 1.2 are no longer supported; this change was necessary for the interleaved parallelism overhaul in this version
  • the interleaved parallelism backend got fixed to properly support changes introduced in Julia 1.2; it now runs more stably and features better error reporting

bug fixes

  • fixed occasional hangs when computing networks in parallel via the interleaved mode (issue #9)

v0.16.0

general

  • add explicit parallelism flag (parallel_mode) to learn_network to avoid having to remove workers when switching from multi-core to single-core computations
  • improve error messages in several places

bug fixes

  • fix edge case in bin filtering: variables w/o zero entries are now retained and properly discretized
  • ensure that meta_mask is a proper BitVector in learn_network and normalize_data

v0.15.0

general

  • improve sign determination for conditional mutual information tests
  • remove jld2 support due to stability issues
  • categorical meta data elements must now include non-numeric characters to ensure dintinguishability from continuous columns

bug fixes

  • fix bug in read_edgelist where files with unconnected nodes at high header positions could not be read
  • fix bug where continuous variables were sometimes one-hot encoded (thanks @pbelmann for reporting)
  • fix bug where zero weights were assigned to highly associated variable pairs in rare cases (thanks @Mortomoto for reporting)
  • fix bug that could lead to the first data column being parsed as row identifiers if it had no duplicate entries
  • fix save_network with detailed=true not producing output in the latest version (thanks @pbelmann for reporting)

v0.14.0

general

  • allow string factors in meta data
  • make one-hot encoding the default for meta data variables with more than two categories, this avoids unreliable tests and improves interpretability (can be switched off with "make_onehot = false")
  • improve support for continuous meta data variables when using "fast=true"
  • support numeric OTU identifiers

bug fixes

  • fix Travis CI osx handling

v0.13.1

  • fix stdlib dependency issue

v0.13.0

general

  • make Travis CI handle osx (fix HDF5 compilation)
  • remove several unmaintained backend options (temporary output, AND rule, pvalue weights)
  • make adaptive CLR normalized matrices always dense to avoid inefficiencies
  • drop samples with adaptive pseudo-counts below machine precision for now

bug fixes

  • fix meta_mask bug in data normalization
  • fix performance regression in CLR transform of sparse matrices