WIP isolate convergence warning in LogReg

.gitignore

@@ -14,7 +14,6 @@ celer.egg-info
 # build
 build
 
-
 # cache
 .pytest_cache
 __pycache__

celer/tests/conv_warning/inspect_conv_warn.py

@@ -0,0 +1,25 @@
+"""For debbuging purposes."""
+
+import numpy as np
+from numpy.linalg import norm
+from sklearn.utils.estimator_checks import check_estimator
+
+from celer.dropin_sklearn import LogisticRegression
+from celer.utils.testing import build_dataset
+
+
+np.random.seed(1409)
+X, y = build_dataset(
+    n_samples=30, n_features=60, sparse_X=True)
+y = np.sign(y)
+alpha_max = norm(X.T.dot(y), ord=np.inf) / 2
+C = 20. / alpha_max
+
+tol = 1e-4
+clf1 = LogisticRegression(C=C, tol=tol, verbose=0)
+
+generator = check_estimator(clf1, generate_only=True)
+generator = list(generator)
+
+for i, (estimator, check_estimator) in enumerate(generator[37:]):
+    check_estimator(estimator)

celer/tests/conv_warning/reproduce_warning.py

@@ -0,0 +1,48 @@
+"""Note:
+
+Use 'celer/tests/conv_warning/logs/dumped_data_.pkl'
+as dumped_data path to run script from terminal.
+"""
+
+import pickle
+import numpy as np
+from numpy.linalg import norm
+from celer import LogisticRegression
+from sklearn.linear_model import LogisticRegression as sk_LR
+
+
+# load data
+dumped_data_path = './logs/dumped_data_.pkl'
+with open(dumped_data_path, 'rb') as f:
+    DICT_DATA = pickle.load(f)
+
+data = DICT_DATA['check_fit_idempotent']
+X = data["X"]  # data centered around 100
+y = data["y"]
+C = data["C"]
+
+C_min = 2 / norm(X.T @ y, ord=np.inf)
+
+# C is very high (higher is more difficult):
+print(f"C / Cmin {C / C_min:.2e}")
+# to get the warning:
+clf = LogisticRegression(C=C, verbose=0).fit(X, y)
+
+# a lower C gets no warning :
+LogisticRegression(C=50 * C_min, verbose=1).fit(X, y)
+
+###############################################################################
+# regular celer converges but takes >= 10_000 epochs:
+clf = LogisticRegression(C=C, solver="celer", verbose=1, tol=1e-10).fit(X, y)
+
+# liblinear does not seem to
+clf_sk = sk_LR(C=C, fit_intercept=False, penalty="l1",
+               solver="liblinear", max_iter=10000, tol=1e-8).fit(X, y)
+
+
+# finally, centering the columns yields convergence in very few iterations
+X_c = X - X.mean(axis=0)
+
+C_min = 2 / np.max(np.abs(X_c.T @ y))
+clf = LogisticRegression(C=C, verbose=2,
+                         solver="celer-pn", tol=1e-10).fit(X_c, y)