Set default step for GLM and l2 penalization to a better value

Useful for comparison with scikit.

Author: Mbompr

tick/base/learner/learner_glm.py

@@ -3,8 +3,8 @@
 from warnings import warn
 
 import numpy as np
-from tick.base import Base
 
+from tick.base import Base
 from tick.base_model import ModelLipschitz
 from .learner_optim import LearnerOptim
 
@@ -167,11 +167,20 @@ def fit(self, X: object, y: np.array):
         if self.step is None and self.solver in self._solvers_with_step:
             if self.solver in self._solvers_with_linesearch:
                 self._solver_obj.linesearch = True
-            elif self.solver == 'svrg':
+            elif self.solver == 'svrg' or self.solver == 'saga':
+                L = self._model_obj.get_lip_max()
+                if self.penalty == 'l2':
+                    L += 1. / self.C
+                    mun = min(2 * self._model_obj.n_samples / self.C, L)
+                    self.step = 1. / (2 * L + mun)
+                else:
+                    self.step = 1. / L
+
                 if isinstance(self._model_obj, ModelLipschitz):
                     self.step = 1. / self._model_obj.get_lip_max()
                 else:
-                    warn('SVRG step needs to be tuned manually', RuntimeWarning)
+                    warn('SVRG and SAGA steps needs to be tuned manually',
+                         RuntimeWarning)
                     self.step = 1.
             elif self.solver == 'sgd':
                 warn('SGD step needs to be tuned manually', RuntimeWarning)