Machine workaround. This would avoid the bindings patch

bob/learn/linear/machine.py

@@ -4,6 +4,7 @@
 
 from ._library import Machine as _Machine_C
 import bob.learn.activation
+import numpy
 
 
 class Machine(_Machine_C):
@@ -22,26 +23,29 @@ class Machine(_Machine_C):
         output_dict["input_sub"] = self.input_subtract
         output_dict["input_div"] = self.input_divide
         output_dict["weights"] = self.weights
-        output_dict["activation"] = self.activation_py.to_dict()["id"]
+        output_dict["activation"] =  bob.learn.activation.Activation.to_dict(self.activation)["id"]
         output_dict["biases"] = self.biases
 
+
+
         return output_dict
 
+
     @classmethod
     def from_dict(cls, input_dict):
         """
         Loads itself from a python dict :py:class:`dict`
         """
 
-        machine = cls(input_dict["weights"])
-        machine.biases = input_dict["biases"]
-        machine.input_subtract = input_dict["input_sub"]
-        machine.input_divide = input_dict["input_div"]
+        machine = cls(numpy.array(input_dict["weights"], dtype="float"))
+        machine.biases = numpy.array(input_dict["biases"], dtype="float")
+        machine.input_subtract = numpy.array(input_dict["input_sub"], dtype="float")
+        machine.input_divide = numpy.array(input_dict["input_div"], dtype="float")
+        
         activation = bob.learn.activation.Activation.from_dict(
             {"id": input_dict["activation"]}
         )
+
         machine.activation = activation
-        # TODO: Temporarily solution to load activation
-        machine.activation_py = activation
 
         return machine