[FIX] address flaky tests in CI (#3785)

* remove duplicate tests

* skip test for now

* only skip test

* silence more tests

* ignore all parcellations tests

* skip one more test

* only run parcellation tests

* refactor tests

* reintroduce test

* reintroduce test

* bring yet another test back in

* switch ubuntu version for doc

* test on all

* use latet

.github/workflows/build-docs.yml

@@ -20,7 +20,7 @@ jobs:
     # This prevents this workflow from running on a fork.
     # To test this workflow on a fork, uncomment the following line.
     if: github.repository == 'nilearn/nilearn'
-    runs-on: ubuntu-20.04
+    runs-on: ubuntu-latest
     timeout-minutes: 360
     env:
       BROWSER: "/usr/bin/firefox"

nilearn/regions/tests/test_parcellations.py

@@ -109,18 +109,6 @@ def test_parcellations_fit_on_multi_nifti_images(method, test_image):
 
     assert parcellator.labels_img_ is not None
 
-    parcellator = Parcellations(method="rena", n_parcels=5, verbose=0)
-    parcellator.fit(fmri_imgs)
-
-    assert parcellator.labels_img_ is not None
-
-    parcellator = Parcellations(
-        method="hierarchical_kmeans", n_parcels=5, verbose=0
-    )
-    parcellator.fit(fmri_imgs)
-
-    assert parcellator.labels_img_ is not None
-
     # Smoke test with explicit mask image
     mask_img = np.ones((10, 11, 12))
     mask_img = Nifti1Image(mask_img, AFFINE_EYE)
@@ -243,20 +231,26 @@ def test_parcellations_transform_with_multi_confounds_multi_images(
 @pytest.mark.parametrize("method", METHODS)
 @pytest.mark.parametrize("n_parcel", [5])
 def test_fit_transform(method, n_parcel, test_image_2):
-    rng = np.random.RandomState(42)
     fmri_imgs = [test_image_2] * 3
-    confounds = rng.standard_normal(size=(10, 3))
-    confounds_list = [confounds] * 3
 
     parcellator = Parcellations(method=method, n_parcels=n_parcel, verbose=0)
-    signals = parcellator.fit_transform(fmri_imgs)
+    parcellator.fit_transform(fmri_imgs)
 
     assert parcellator.labels_img_ is not None
     if method not in ["kmeans", "rena", "hierarchical_kmeans"]:
         assert parcellator.connectivity_ is not None
     assert parcellator.masker_ is not None
 
-    # fit_transform with confounds
+
+@pytest.mark.parametrize("method", METHODS)
+@pytest.mark.parametrize("n_parcel", [5])
+def test_fit_transform_with_condounds(method, n_parcel, test_image_2):
+    rng = np.random.RandomState(42)
+    fmri_imgs = [test_image_2] * 3
+    confounds = rng.standard_normal(size=(10, 3))
+    confounds_list = [confounds] * 3
+
+    parcellator = Parcellations(method=method, n_parcels=n_parcel, verbose=0)
     signals = parcellator.fit_transform(fmri_imgs, confounds=confounds_list)
 
     assert isinstance(signals, list)