Merge pull request #510 from bashtage/enable-cow-support

MAINT: Enable copy-on-write support

Author: bashtage

ci/azure_template_posix.yml

@@ -59,6 +59,10 @@ jobs:
       python311_latest:
         python.version: '3.11'
         XXHASH: true
+      python311_copy_on_write:
+        python.version: '3.11'
+        XXHASH: true
+        LM_TEST_COPY_ON_WRITE: 1
     maxParallel: 10
 
   steps:

linearmodels/conftest.py

@@ -1,7 +1,21 @@
 from __future__ import annotations
 
+import logging
+import os
+
+import pandas as pd
 import pytest
 
+logger = logging.getLogger(__name__)
+
+
+try:
+    cow = bool(os.environ.get("LM_TEST_COPY_ON_WRITE", False))
+    pd.options.mode.copy_on_write = cow
+    logger.critical("Copy on Write Enabled!")
+except AttributeError:
+    logger.critical("Copy on Write disabled")
+
 
 def pytest_configure(config):
     # Minimal config to simplify running tests from lm.test()

linearmodels/iv/absorbing.py

@@ -22,6 +22,7 @@
     ndarray,
     ones,
     ptp,
+    require,
     sqrt,
     where,
     zeros,
@@ -702,7 +703,7 @@ def __init__(
         self._regressors_hash: tuple[tuple[str, ...], ...] | None = None
 
     def _drop_missing(self) -> BoolArray:
-        missing = self.dependent.isnull.to_numpy()
+        missing = require(self.dependent.isnull.to_numpy(), requirements="W")
         missing |= self.exog.isnull.to_numpy()
         missing |= self._absorb_inter.cat.isnull().any(axis=1).to_numpy()
         missing |= self._absorb_inter.cont.isnull().any(axis=1).to_numpy()

linearmodels/panel/data.py

@@ -284,7 +284,7 @@ def dataframe(self) -> DataFrame:
     @property
     def values2d(self) -> AnyArray:
         """NumPy ndarray view of dataframe"""
-        return np.asarray(self._frame)
+        return np.require(self._frame, requirements="W")
 
     @property
     def values3d(self) -> AnyArray:
@@ -502,7 +502,7 @@ def demean_pass(
 
         exclude = np.ptp(np.asarray(self._frame), 0) == 0
         max_rmse = np.sqrt(np.asarray(self._frame).var(0).max())
-        scale = np.asarray(self._frame.std())
+        scale = np.require(self._frame.std(), requirements="W")
         exclude = exclude | (scale < 1e-14 * max_rmse)
         replacement = cast(Float64Array, np.maximum(scale, 1))
         scale[exclude] = replacement[exclude]

linearmodels/panel/model.py

@@ -2419,7 +2419,7 @@ def fit(
             )
             w_frame = w_frame.reindex(self.weights.index).dropna(how="any")
             index = cast(MultiIndex, w_frame.index)
-            w = w_frame.to_numpy()
+            w = np.require(w_frame, requirements="W")
 
             w /= w.mean()
             root_w = cast(Float64Array, np.sqrt(w))

linearmodels/tests/panel/test_data.py

@@ -295,7 +295,9 @@ def test_demean_against_dummy_regression(data):
 
 
 def test_demean_missing(mi_df):
-    mi_df.values.flat[::13] = np.nan
+    values = np.require(mi_df, requirements="W")
+    values.flat[::13] = np.nan
+    mi_df.loc[:, :] = values
     data = PanelData(mi_df)
     fe = data.demean("entity")
     expected = data.values3d.copy()

linearmodels/tests/panel/test_model.py

@@ -198,8 +198,8 @@ def test_absorbing_effect(data, intercept):
         nentity = len(x.index.levels[0])
         ntime = len(x.index.levels[1])
         temp = data.x.iloc[:, 0].copy()
-        temp.values[:] = 1.0
-        temp.values[: (ntime * (nentity // 2))] = 0
+        temp.loc[:] = 1.0
+        temp.iloc[: (ntime * (nentity // 2))] = 0
 
         if intercept:
             x["Intercept"] = 1.0