Towards Explanation of DNN-based Prediction with Guided Feature Inversion