play nicer with discontiguous inputs, skl_groups Features objects

mmd/mmd.py

@@ -104,18 +104,26 @@ def _atonce_rbf_mmk(X, Y=None, gammas=1, get_X_diag=False, get_Y_diag=False):
         return K[0] if scalar_gamma else K
 
 
+def _get_stacked(X):
+    if hasattr(X, 'n_samps'):  # is an skl_groups Features object
+        X.make_stacked()
+        return X.stacked_features, X.n_samps
+
+    X_stacked = np.ascontiguousarray(np.vstack(X))
+    X_n_samps = np.array([len(x) for x in X], dtype=np.int32)
+    return X_stacked, X_n_samps
+
+
 def _el_by_el_rbf_mmk(X, Y=None, gammas=1, get_X_diag=False, get_Y_diag=False,
                       n_jobs=1):
-    X_stacked = np.vstack(X)
-    X_n_samps = np.array([len(x) for x in X], dtype=np.int32)
+    X_stacked, X_n_samps = _get_stacked(X)
 
     if Y is None or Y is X:
         Y_stacked = X_stacked
         Y_n_samps = X_n_samps
         X_is_Y = True
     else:
-        Y_stacked = np.vstack(Y)
-        Y_n_samps = np.array([len(y) for y in Y], dtype=np.int32)
+        Y_stacked, Y_n_samps = _get_stacked(Y)
         X_is_Y = False
 
     if all(isinstance(h, logging.NullHandler)