MNT Simply gradient computation of sigmoid calibration (scikit-learn#14439)

Author: qinhanmin2014

sklearn/calibration.py

@@ -452,9 +452,8 @@ def objective(AB):
 
     def grad(AB):
         # gradient of the objective function
-        E = np.exp(AB[0] * F + AB[1])
-        P = 1. / (1. + E)
-        TEP_minus_T1P = P * (T * E - T1)
+        P = expit(-(AB[0] * F + AB[1]))
+        TEP_minus_T1P = T - P
         if sample_weight is not None:
             TEP_minus_T1P *= sample_weight
         dA = np.dot(TEP_minus_T1P, F)