[MRG] BUG Fixes test_scale_and_stability in windows (#15661)

thomasjpfan · jnothman · commit b3188fbe2d30 · 2019-11-28T20:51:43.000+11:00
diff --git a/doc/whats_new/v0.22.rst b/doc/whats_new/v0.22.rst
@@ -101,6 +101,7 @@ random sampling procedures.
 - :class:`linear_model.Ridge` when `X` is sparse. |Fix|
 - :class:`model_selection.StratifiedKFold` and any use of `cv=int` with a
   classifier. |Fix|
+- :class:`cross_decomposition.CCA` when using scipy >= 1.3 |Fix|
 
 Details are listed in the changelog below.
 
@@ -209,6 +210,9 @@ Changelog
   ``inverse_transform`` to transform data to the original space`.
   :pr:`15304` by :user:`Jaime Ferrando Huertas <jiwidi>`.
 
+- |Fix| :class:`cross_decomposition.CCA` now produces the same results with 
+  scipy 1.3 and previous scipy versions. :pr:`15661` by `Thomas Fan`_.
+
 :mod:`sklearn.datasets`
 .......................
 
diff --git a/sklearn/cross_decomposition/_pls.py b/sklearn/cross_decomposition/_pls.py
@@ -40,14 +40,26 @@ def _nipals_twoblocks_inner_loop(X, Y, mode="A", max_iter=500, tol=1e-06,
     ite = 1
     X_pinv = Y_pinv = None
     eps = np.finfo(X.dtype).eps
+
+    if mode == "B":
+        # Uses condition from scipy<1.3 in pinv2 which was changed in
+        # https://github.com/scipy/scipy/pull/10067. In scipy 1.3, the
+        # condition was changed to depend on the largest singular value
+        X_t = X.dtype.char.lower()
+        Y_t = Y.dtype.char.lower()
+        factor = {'f': 1E3, 'd': 1E6}
+
+        cond_X = factor[X_t] * eps
+        cond_Y = factor[Y_t] * eps
+
     # Inner loop of the Wold algo.
     while True:
         # 1.1 Update u: the X weights
         if mode == "B":
             if X_pinv is None:
                 # We use slower pinv2 (same as np.linalg.pinv) for stability
                 # reasons
-                X_pinv = pinv2(X, check_finite=False)
+                X_pinv = pinv2(X, check_finite=False, cond=cond_X)
             x_weights = np.dot(X_pinv, y_score)
         else:  # mode A
             # Mode A regress each X column on y_score
@@ -64,7 +76,8 @@ def _nipals_twoblocks_inner_loop(X, Y, mode="A", max_iter=500, tol=1e-06,
         # 2.1 Update y_weights
         if mode == "B":
             if Y_pinv is None:
-                Y_pinv = pinv2(Y, check_finite=False)  # compute once pinv(Y)
+                # compute once pinv(Y)
+                Y_pinv = pinv2(Y, check_finite=False, cond=cond_Y)
             y_weights = np.dot(Y_pinv, x_score)
         else:
             # Mode A regress each Y column on x_score
