Added linear scaling to ScaleOnPlateau

keras.py

@@ -488,7 +488,7 @@ class PatientTracker(BestTracker):
 
 
 class ScaleOnPlateau(PatientTracker):
-    def __init__(self, target, factor, min=None, max=None, verbose=0, log_key=None, **kwargs):
+    def __init__(self, target, factor, min=None, max=None, verbose=0, log_key=None, linearly=False, **kwargs):
         pin(locals(), kwargs)
         super(ScaleOnPlateau, self).__init__(**kwargs)
 
@@ -509,7 +509,10 @@ class ScaleOnPlateau(PatientTracker):
         if self.log_key is not None:
             logs.setdefault(self.log_key, cur)
         if self.patient_step(epoch, logs) == "good":
-            new = cur * self.factor
+            if self.linearly:
+                new = cur + self.factor
+            else:
+                new = cur * self.factor
             if self.min is not None:
                 new = max(new, self.min)
             if self.max is not None: