Skip to content

Latest commit

 

History

History
209 lines (138 loc) · 6.46 KB

changelog.md

File metadata and controls

209 lines (138 loc) · 6.46 KB

1.4.0

Interface changes

  • Add --pimp_interactive-flag to toggle bokeh-plotting for pimp-plots
  • Add --pimp_whiskers-flag to toggle plotting of pimp-whiskers plot

Major changes

  • Add basic Auto-PyTorch file format support
  • Add tensorboard support for notebooks
  • Add custom APT-overview tables
  • Use interactive bokeh plots for pimp

Minor changes

  • Add lots of docstring-documentation, PEP8 and a contribute-guide
  • Rename some classes and functions to unify naming (only internal)
  • Fix naming convention for hpbanster conversion folders

1.3.3

Interface changes

  • Deprecate deactivation of individual analysis methods using the --no_[analysis_name] flag
  • Add two mutually exclusive flags --[skip|only] [analysis_name] to activate / skip specific analysis methods
  • Remove --pimp_sort_table_by flag (table is sortable interactively now so no need)

Major changes

  • Drop support for Python 3.5, Add support for Python 3.7 and 3.8
  • Drop support for SMAC 0.12.0, Add support for SMAC 0.12.1 and 0.12.2
  • Distinguish between per-parallel-run and per-budget meta information (#270)

Minor changes

  • Change name of scatter plot from "Scatterplot" to "Scatter Plot"

1.3.2

Bug fixes

  • Fix bug with duplicate entries in html-builder

Major changes

  • Internally adapt structure to support budgets and parallel runs in accordance with new SMAC
  • Enable budgets and parallel runs for CSV-Reader (and change to converting data always (so also in hpbandster-reader) instead of maintaining different readers)
  • Add interactive parallel coordinates plot
  • Add Whisker-Quantile plot to compare importances over budgets and parallel runs

Minor changes

  • Add support for new trajectory files (overdue...)
  • Add script to enable generation of example-data in CSV-format
  • Add caching of aggregated ConfiguratorRuns
  • Update CSV-examples (now with budgets instead of instances...)

1.3.1

Bug fixes

  • Fix #261, get_checkbox threw an error due to unpredictable return type

1.3.0

Major changes

  • Support SMAC 0.12.0
  • Drop support for SMAC < 0.12.0 (use CAVE 1.2.x instead)

Minor changes

  • PyImp min version now 1.0.8

1.2.0

Bug fixes

  • Fix tooltips sometimes not showing
  • Fix docu-build
  • Fix logging (was printing sometimes)
  • Fix bug in reading in configuration in CSV-data

Major changes

  • Change internal structure from cave-in-cave to proper data-handling using a RunsContainer with individual ConfiguratorRuns (per budget/parallel-run combination)
  • Add options-file in .ini format to increase flexibility on default options (and technically enable using options-file for CAVE, though not supported by cmdline yet, only in python/jupyter)
  • Make the --folders-cmd line argument positional (though the keyword --folders is still supported)
  • Enable automatic file-format detection (--file_format-cmd line argument still supported)

Minor changes

  • Add whitebox-tests and script to execute all examples
  • Add exceptions for Deactivated and NotApplicable errors and load a lot more work onto BaseAnalyzer

1.1.8

Bug fixes

  • Fix args-Errors when trying to use with SMAC 0.11.x
  • Fix overflow in blcs checkboxes

Minor changes

  • Force pyimp >= 1.0.6

1.1.7

Bug fixes

  • Fix error's in some plots related to NaN's in bohb-trajectory
  • Fix fanova display error in jupyter notebook
  • Fix critical error on rh-reduction in cfp (individual budget's didn't reduce, leading to KeyError)
  • Fix too many cfp-plots (for individual budgets) in some cases
  • Fix several jupyter-related bugs
  • (minor) Fix TypeError when additional-info in bohb-result is a list (not a dict)

Major changes

  • Add option cot_inc_traj to choose the way the bohb-trajectory is interpreted (add racing method)
  • Add ConfigSpace to overview-table
  • Budgets with long floats are rounded as much as possible while staying unique
  • Add hpbandster to requirements (overdue, this is mainly used for hpbandster)

Minor changes

  • Update bokeh to 1.1.0
  • Format issues (dos to unix)

1.1.6

Bug fixes

  • Recreate output always, zip old files if present
  • Fix jupyter-notebook logging and output-folder recreation

Minor changes

  • Show budgets as integers, if they are integers
  • Add single budgets to cost-over-time plot (from bohb-result)
  • Add std to bohb-runs in cost-over-time plot
  • Remove combined option from contour-plot-selection, fix checkboxes
  • Improve budget correlation plot (add dropdown-selection)
  • Remove package orange3

1.1.5

Bug fixes

  • Parallel Coordinates order now derived from parameter importance
  • Show all fANOVA pairwise plots
  • Fix bug in cost-over-time if run multiple times with same output-dir
  • Fix multiple bugs related to floats in Categoricals
  • Fix surface prediction in cfp when forbidden conditions
  • Fix path-spec error on loading SMAC-data
  • Sub-dict's were not generated in HTML-report sometimes
  • Parameters active in either def or inc are shown in comparison (before only active in def)

Minor changes

  • Add number of bohb-runs to meta-table
  • Make overflowing HTML-tables scrollable
  • Don't plot scatter or cdf if leq than one instance
  • Light changes in output-dir-structure
  • Evaluate all parameters in importance
  • Group cmd-line-args
  • 'Cost' instead of 'Quality' label in all plots
  • Redesign BOHB-lcs-plot to allow for 100+ iterations

1.1.4

Bug fixes

  • Enforcing smac version 0.10.0 to fix pcs-loading issues for booleans
  • Catch NaN-values from BOHB-results
  • Fix sorting-bug in pimp-table
  • Fix notebook-errors when executing cfp

Major changes

  • New version of the overview-table (with run-specific for every run)
  • Enable support for multiple BOHB-runs (just aggregating budgets across results)

Minor changes

  • Make nomenclature for BOHB clearer
  • Add BOHB to extensive-testing-suite

1.1.1

Major changes

  • Change support for BOHB-reports, creating one html-file with all budgets as sub-reports now.
  • There are now two ways to generate a report: one uses the given runs as equal, aggregates them and trains an EPM on all of them, the other is the use-budgets option that tries to treat them as separate runs, that can only be compared in certain fashion
  • Fix critical errors in csv- and smac2- reader
  • Catch failed bokeh-exports
  • Add algorithm footprints mouseover
  • Add fanova std's when using bleeding edge of pyimp and fanova

Interface changes

  • passing most of the relevant arguments directly from facade to analyzer now (not during analyzer's init)

Minor changes

  • improving documentation and comments, add custom logos