Structured Term Pruning for Computational Efficient Neural Networks Inference