support max_iter = 0 for Logreg with prox newton solver (#213)

Author: mathurinm

celer/PN_logreg.pyx

@@ -41,9 +41,10 @@ def newton_celer(
     # scale tol for when problem has large or small p_obj
     tol *= n_samples * np.log(2)
 
-    cdef int i, j, t, k
-    cdef floating p_obj, d_obj, gap, norm_Xtheta, norm_Xtheta_acc
-    cdef floating tmp
+    cdef int t = 0
+    cdef int i, j, k
+    cdef floating p_obj, d_obj, norm_Xtheta, norm_Xtheta_acc, tmp
+    cdef floating gap = -1  # initialized for the warning if max_iter=0
     cdef int info_dposv
     cdef int ws_size
     cdef floating eps_inner = 0.1
@@ -54,7 +55,8 @@ def newton_celer(
     cdef int[:] all_features = np.arange(n_features, dtype=np.int32)
     cdef floating[:] prios = np.empty(n_features, dtype=dtype)
     cdef int[:] WS
-    cdef floating[:] gaps = np.zeros(max_iter, dtype=dtype)
+    cdef floating[:] gaps = np.zeros(max(1, max_iter), dtype=dtype)
+    gaps[0] = gap  # support max_iter = 0
     cdef floating[:] X_mean = np.zeros(n_features, dtype=dtype)
     cdef bint center = False
     # TODO support centering

celer/lasso_fast.pyx

@@ -59,7 +59,7 @@ def celer(
     cdef int i, j, k, idx, startptr, endptr, epoch
     cdef int ws_size = 0
     cdef int nnz = 0
-    cdef floating gap = -1
+    cdef floating gap = -1  # initialized for the warning if max_iter=0
     cdef floating p_obj, d_obj, highest_d_obj, radius, tol_in
     cdef floating gap_in, p_obj_in, d_obj_in, d_obj_accel, highest_d_obj_in
     cdef floating tmp, R_sum, tmp_exp, scal

celer/tests/test_lasso.py

@@ -16,7 +16,7 @@
                                   Lasso as sklearn_Lasso, lasso_path)
 
 from celer import celer_path
-from celer.dropin_sklearn import Lasso, LassoCV
+from celer.dropin_sklearn import Lasso, LassoCV, LogisticRegression
 from celer.utils.testing import build_dataset
 
 
@@ -217,5 +217,20 @@ def test_weights():
     np.testing.assert_raises(ValueError, clf1.fit, X=X, y=y)
 
 
+def test_zero_iter():
+    X, y = build_dataset(n_samples=30, n_features=50)
+
+    # convergence warning is raised bc we return -1 as gap
+    with warnings.catch_warnings(record=True):
+        assert_allclose(Lasso(max_iter=0).fit(X, y).coef_, 0)
+        y = 2 * (y > 0) - 1
+        assert_allclose(
+            LogisticRegression(max_iter=0, solver="celer-pn").fit(X, y).coef_,
+            0)
+        assert_allclose(
+            LogisticRegression(max_iter=0, solver="celer").fit(X, y).coef_,
+            0)
+
+
 if __name__ == "__main__":
     pass