Signed-off-by: MichaelTiemann <[email protected]>

Big remaining problem is unhanshable UFloat type.

Signed-off-by: MichaelTiemann <[email protected]>

Signed-off-by: [email protected]

Signed-off-by: MichaelTiemann <[email protected]>

Signed-off-by: MichaelTiemann <[email protected]>

Big remaining problem is unhanshable UFloat type.

Signed-off-by: MichaelTiemann <[email protected]>

Signed-off-by: [email protected]

Signed-off-by: MichaelTiemann <[email protected]>

…C/pint-pandas into ducks-unlimited

Signed-off-by: Michael Tiemann <[email protected]>

Deal with duality of np.nan and _ufloat_nans when constructing PintArrays, creating float64 magnitude arrays when possible and ufloat magnitudes when necessary.  Note that PintArray columns don't advertise whether magnitudes are np.float64 or np.obejct; they are what they are.  Also commented out warning, which simply creates noise.

Signed-off-by: Michael Tiemann <[email protected]>

Fix conditional expression.

Signed-off-by: Michael Tiemann <[email protected]>

Don't try to build a large-enough-to-hold-uncertainties array if we are not using uncertainties in any way.

Signed-off-by: Michael Tiemann <[email protected]>

The manual merge process created a syntax error and various things that black didn't like.

Signed-off-by: Michael Tiemann <[email protected]>

Signed-off-by: Michael Tiemann <[email protected]>

Fixes include:
* Factorization
* NaN handling (several more issues still need to be resolved)
* Proper unit declarations in test_offset_concat
* Integration of new `numeric_dtype` parameter

A major outstanding issue (presently being discussed as pandas-dev/pandas#53904) concerns whether we can make AffineScalarFunc hashable and/or whether other legacy Pandas code (which has been deprecated) can be further removed.

Signed-off-by: Michael Tiemann <[email protected]>

Make `ruff` and `black` happy.

Signed-off-by: Michael Tiemann <[email protected]>

To resolve the question of the proper na_value for EA dtypes (np.nan vs. uncertainties _ufloat_nan), use the gender-neutral pd.NA value.

Signed-off-by: Michael Tiemann <[email protected]>

Make black happy.

Signed-off-by: Michael Tiemann <[email protected]>

With these changes (and pandas-dev/pandas#53970 and hgrecco/pint#1615) the test suite passes or xpasses everything (no failures or error).  Indeed, the uncertainties code has essentially doubled the scope of the test suite (to test with and without it).

The biggest gotcha is that the EA for complex numbers is not compatible with the EA for uncertainties, due to incompatible hacks:

The hack for complex numbers is to np.nan (which is, technically, a complex number) for na_value across all numeric types.  But that doesn't work for uncertainties, because uncertainties doesn't accept np.nan as an uncertain value.

The hack for uncertainties is to use pd.NA for na_value.  This works for Int64, Float64, and uncertainties, but doesn't work for complex (which cannot tolerate NAType).

Some careful subclassing fills in what doesn't easily work, with fixtures to prevent the improper mixing of complex and uncertainty types in the same python environment.  Happy to discuss!

Signed-off-by: Michael Tiemann <[email protected]>

Except that ruff cannot be made happy, see astral-sh/ruff#2044

Signed-off-by: Michael Tiemann <[email protected]>

Moving the # noqa: F811 to the correct line allows the file to pass ruff.

Signed-off-by: Michael Tiemann <[email protected]>

Arrgh!

Signed-off-by: Michael Tiemann <[email protected]>

Missed this.

Signed-off-by: Michael Tiemann <[email protected]>

Issue hgrecco#174 reports that DataFrame reduction was broken by the latest Pint-Pandas changes.  This commit adapts Pint-Pandas to work with upcoming Pandas 2.1, currently scheduled for release Aug 20, 2023.

Signed-off-by: Michael Tiemann <[email protected]>

Signed-off-by: Michael Tiemann <[email protected]>

Remove print statement mistakenly copied into _accumulate code that was never (properly) tested.

Signed-off-by: Michael Tiemann <[email protected]>

Using np.nan actually simplifies the integration of uncertainties into pint_pandas.  Recent changes in Pandas 2.1 are now sufficient to pass all Pint-Pandas tests.

Also added a fixture validating that we can use either `np.nan` or `ufloat(np.nan, 0)` as NA value (the latter being something that should naturally arise in the course of uncertainties calculations, not just missing data from the start).

Signed-off-by: Michael Tiemann <[email protected]>

The internal Pandas 2.1 interfaces continue to move and change.  These changes adapt to those changes, adjsting and removing some redundant subclass methods.

Signed-off-by: Michael Tiemann <[email protected]>

Merge in important bugfixes for `na_value` handling.  And keep `black` happy.

Signed-off-by: Michael Tiemann <[email protected]>

First draft of 100% working test cases after (re)merging with changes extracted from these changes as part of PR hgrecco#196.

Signed-off-by: Michael Tiemann <[email protected]>

Fix so that it works in a world without uncertainties (or with uncertainties).

Signed-off-by: Michael Tiemann <[email protected]>

Signed-off-by: Michael Tiemann <[email protected]>

For now disable testing of `uncertainties` specifically, as that also requires an uncertainties-aware `pint`.  When that becomes available we can update the CI/CD scripts to enable optional uncertainty testing with versions of pint and pint-pandas that can tolerate it.

From local testing with uncertainties, remember: if you don't put a fixture XYZ in the parameter list, then testing `if XYZ` is really testing whether the fixture's function is not None (it's always not None).  When XYZ is in the parameter list, testing `if XYZ` tells you whether the value of XYZ fixture is non-False.

Also increase version compatibility by using `np.array([f(x) for x in arr])` instead of `arr.map(f)` when `arr` is a numpy array.

Signed-off-by: Michael Tiemann <[email protected]>

Added blurb about supporting `uncertainties`.

Signed-off-by: Michael Tiemann <[email protected]>

Settle changes that work with Pandas 2.1.0 and Pint-0.23rc0 and see if we can pass Pint-Pandas CI/CD.

Signed-off-by: Michael Tiemann <[email protected]>

Allow pip install command to unpack uncertainties option into multiple arguments.

Signed-off-by: Michael Tiemann <[email protected]>

By using `include` directive, we can add a special case to the matrix of possibilities.  In this case we add to the matrix the case where pandas>=2.1.0 and pint>=0.23rc0, which are both necessary for the UFloat array changes to work.

Signed-off-by: Michael Tiemann <[email protected]>

Don't test uncertainties in ci-pint-pre or ci-pint-master, as these don't have the necessary versions of Pandas or Pint to make the matrix include operation work.

Signed-off-by: Michael Tiemann <[email protected]>

Also test Pint-Pandas without uncertainties.

Signed-off-by: Michael Tiemann <[email protected]>

Trying again with different syntax to avoid duplicate key issue.

Signed-off-by: Michael Tiemann <[email protected]>

Trying yet another syntax to make `uncertainties` one of two options when pandas==2.1.0 and pint==0.23rc0.

Signed-off-by: Michael Tiemann <[email protected]>

Try using null instead of "" for uncertainties that we don't want to run.

Signed-off-by: Michael Tiemann <[email protected]>

Try using "null" to represent a condition where we don't want uncertainties installed.

Signed-off-by: Michael Tiemann <[email protected]>

Try using "no-thank-you" instead of "null" for matrix permutations not intended to use uncertainties.

Signed-off-by: Michael Tiemann <[email protected]>

Wrap `uncertainties` matrix parameter in `[]`.  If we get this working, can try setting parameter to `''`' from several commits ago.

Signed-off-by: Michael Tiemann <[email protected]>

Use single quotes when testing the value of `uncertainties`.

Signed-off-by: Michael Tiemann <[email protected]>

Tweak conditional matrix logic by using a "same-but-different" value for `pint` value assignment.

Signed-off-by: Michael Tiemann <[email protected]>

Untabify yaml file...

Signed-off-by: Michael Tiemann <[email protected]>

Make uncertainties a default null value in the matrix.

Signed-off-by: Michael Tiemann <[email protected]>

Attempt to explicitly add python-version and numpy to `uncertainties` case.  Otherwise, pytest never gets installed.

Signed-off-by: Michael Tiemann <[email protected]>

We cannot reference the matrix when building the matrix, so unroll all parameter values.

Signed-off-by: Michael Tiemann <[email protected]>

Remove unused `uncertainties` from -pre and -master CI files.

Only test one configuration of Python against uncertainties (trying to work around duplicate key problem).

Signed-off-by: Michael Tiemann <[email protected]>