Fix randomized PCA transform after fit

dask_ml/decomposition/pca.py

@@ -193,6 +193,9 @@ def __init__(
         self.svd_solver = svd_solver
         self.tol = tol
         self.iterated_power = iterated_power
+        # scikit-learn's PCA.__sklearn_tags__ reads this attribute for
+        # randomized solvers when check_is_fitted calls get_tags.
+        self.power_iteration_normalizer = "auto"
         self.random_state = random_state
 
     def fit(self, X, y=None):

tests/test_pca.py

@@ -104,6 +104,17 @@ def test_pca_randomized_solver():
     )
 
 
+def test_pca_randomized_transform_after_fit():
+    pca = dd.PCA(n_components=2, svd_solver="randomized", random_state=0)
+
+    pca.fit(dX)
+    assert pca.power_iteration_normalizer == sd.PCA().power_iteration_normalizer
+
+    X_r = pca.transform(dX)
+    assert X_r.shape == (n_samples, 2)
+    X_r.compute()
+
+
 def test_no_empty_slice_warning():
     if not DASK_2_26_0:
         # See https://github.com/dask/dask/pull/6591