Created by: pengwangucla

Current v2 api has evaluator inside of cost layer :

def classification_cost(input,
                        label,
                        weight=None,
                        name=None,
                        top_k=None,
                        evaluator=classification_error_evaluator,
                        layer_attr=None):

This is really wired, why cost layer have such an input, so one must implement a specified type for evaluation.

It is better that evaluator is also a layer, and user can configure outside and connect as a part in the graph. It will be called during test while not called in training.

@reyoung