An Energy-Efficient Inference Method in Convolutional Neural Networks Based on Dynamic Adjustment of the Pruning Level.