[alejandroschuler]

Closes #168.

Problem

For LogNormal, dist.params uses the names from scipy.stats.lognorm: params['s'] is σ and params['scale'] is exp(μ), where μ and σ are the mean and standard deviation of log(Y). The attributes dist.loc and dist.scale held μ and σ. So dist.scale and dist.params['scale'] had the same name but different values. Also, dist.s returned None without an error, because the __getattr__ of the distributions built on scipy.stats returned None for every name that the scipy object does not have.

Changes

• LogNormal has new attributes mu and sigma. loc and scale are now properties. They still return μ and σ, but they give a FutureWarning that names the replacement. The LogNormal scores use mu and sigma, so fitting gives no warning. params does not change.
• Distn has one __getattr__. It forwards a name to self.dist and raises AttributeError for all other names. It replaces the copies in Normal, LogNormal, Exponential, Gamma, HalfNormal, Laplace, Poisson, the three T classes and Weibull. It reads dist from __dict__, so pickle.loads and copy.deepcopy of a predicted distribution now work. Before, both raised RecursionError.
• Docs: the user guide explains the LogNormal params keys and the new attributes. The developer guide no longer tells authors to write a __getattr__ that returns None.
• tests/test_distn_attributes.py tests these changes. 24 of its 26 tests fail on master.

Why not the 2020 plan

The issue thread proposed a leading underscore on the attributes of all distributions. This PR does not do that:

• For the other distributions, the attributes and params agree, and a lot of code reads .loc and .scale on Normal. Examples are NREL's routee-powertrain and this repo's figures/.
• The SymPy factory from SymPy-powered distribution factory for NGBoost #403 names the attributes after the parameter symbols and passes them to user sample_fn and mean_fn. The example notebooks also read the attributes.

This PR also does not change LogNormal params. skpro's NGBoostRegressor and NGBoostSurvival compute μ as log(params['scale']) and σ as params['s']. A change in meaning would give their users wrong numbers without an error.

Compatibility

• Pickled models: a model stores the distribution class and the internal parameters (μ, log σ), not distribution objects. Models pickled on master load on this branch with identical predictions and params, and they give no warning. I checked this for LogNormal regression, LogNormal survival, Exponential survival and Normal.
• skpro 2.16 NGBoostRegressor(dist="LogNormal") gives the same means as ngboost on this branch.
• hasattr(dist, name) is now False for a name that the distribution does not have. Before, it was always True.
• This PR does not set a release for the removal of loc and scale. The warning says "a future release".

Tests

• pytest tests --slow -o "filterwarnings=error::FutureWarning:ngboost" on 6282515: 190 passed, 1 skipped, 2 xfailed, 22 xpassed. No FutureWarning comes from ngboost code, so nothing in ngboost reads LogNormal.loc or LogNormal.scale.
  • The skip is the skops test in tests/test_pickling.py, because skops was not installed. With skops installed, that file gives 4 passed.
  • The 2 xfails are the T cases of test_dists_runs_on_examples_logscore. The 22 xpasses are the other cases of that test. They come from the existing xfail at tests/test_distns.py:91. Its condition is a function, which is always true.
• Why -o and not -W: an earlier version of this description gave pytest tests -W error::FutureWarning:ngboost. That filter does not do this check. On the command line, pytest escapes the module name and matches it exactly. So it catches only warnings from the module ngboost, not from ngboost.distns.lognormal. In the ini option filterwarnings, the module is a regex that matches at the start of the name, so it also catches the submodules. To test the filter, I put one read of self.scale in the LogNormal LogScore metric in a scratch copy. With -W, test_survival_runs_on_examples passed with 500 warnings. With -o, it failed.
• black, isort, flake8 and pylint, as in .pre-commit-config.yaml.

Overlap

#431 also edits docs/content/1-useage.ipynb and docs/content/5-dev.ipynb. This PR changes source lines in two cells of each notebook, so a conflict, if any, will be small.

🤖 Generated with Claude Code

[ryan-wolbeck]

will need to fix some merge conflicts with 431 but looks good

[alejandroschuler]

Thanks! The conflicts with #431 are fixed. f78949e merges master (with #431) into this branch, and the docs build, which runs the notebooks, passes.

I added two small commits so that this PR and #434 merge cleanly together:

• 1f4608b moves the AttributeError release note to a different line of the Bug fixes list. Stop overflow warnings from rejected line-search steps #434 adds its notes after the same line, so the two PRs conflicted.
• 5cdb5e7 disables pylint duplicate-code in the tests lint hook. pre-commit splits the test files over parallel pylint runs. With the new test files of Add mu and sigma to LogNormal and raise AttributeError for unknown names #433 and Stop overflow warnings from rejected line-search steps #434, test_helpers.py and test_pickling.py land in the same run on a 4-CPU runner, and their existing duplicate block fails the hook.