Skip to content

1.5.0

Latest

Choose a tag to compare

@reidjohnson reidjohnson released this 07 Nov 01:52
· 1 commit to 1.5.X since this release

This version adds support for Python version 3.15 and for missing values, and adds wheels for arm64 Linux. It drops support for Python 3.9 and 3.10 and now requires scikit-learn 1.9 or later. The leaf samples are now weighted by their sample weights, and the default interpolation is now "averaged_inverted_cdf", which matches scikit-learn's quantile definition, so predictions can change. The training samples of each leaf are now stored without padding, which can greatly reduce the memory taken by a model and its pickle. The sparse_pickle parameter of fit is removed. Models pickled by earlier versions still load, but earlier versions cannot load models pickled by this version.

What's Changed

Breaking changes

  • Weight leaf samples by their sample weights, and default to scikit-learn's quantile definition by @reidjohnson in #190
  • Store the training samples of each leaf without padding by @reidjohnson in #201
  • Prepare 1.5.0: release notes, nightly scikit-learn tests, cleanups by @reidjohnson in #202

New features

Performance

  • Make weighted predictions and proximity counts independent of training size by @reidjohnson in #179
  • Speed up out-of-bag sample lookups by @reidjohnson in #178
  • Store training targets as an array, so predict doesn't copy them by @reidjohnson in #184
  • Compute OOB quantiles without a Python loop over rows by @reidjohnson in #185
  • Compute monotonic bounds one tree level at a time by @reidjohnson in #186

Bug fixes

  • Avoid changing global random state during leaf subsampling by @Ra1nhy in #167
  • Fix sample weight indexing when collecting leaf samples by @AryanRG in #168
  • Fix float max_samples_leaf in predict and leaf padding by @reidjohnson in #170
  • Stop quantile_ranks from resetting fitted feature attributes by @reidjohnson in #172
  • Validate sample weights before storing them in fit by @reidjohnson in #174
  • Fix proximity_counts argument checks and return types, and docstrings by @reidjohnson in #177
  • Fix models loaded from sparse pickles or read-only memory maps by @reidjohnson in #180
  • Fix sparse inputs in predict by @reidjohnson in #173
  • Validate quantiles as array-likes and reject NaN or empty quantiles by @reidjohnson in #175
  • Declare that missing values are not supported, and reject them up front by @reidjohnson in #176
  • Fix weighted quantiles that use zero-weight inputs or miss exact ranks by @reidjohnson in #171
  • Group leaf samples with a stable sort, so results don't depend on the CPU by @reidjohnson in #181
  • Validate inputs once for OOB samples, without a warning per tree by @reidjohnson in #182
  • Use the fitted trees, not n_estimators, in predict by @reidjohnson in #183
  • Rank among the clipped leaf values in quantile_ranks with monotonic constraints by @reidjohnson in #188
  • Fix small issues found in a review of main by @reidjohnson in #199
  • Redraw OOB samples as the forest drew them, for older models by @reidjohnson in #200

Documentation

Maintenance

New Contributors

Full Changelog: v1.4.2...v1.5.0