A Speed- and Energy-Driven Holistic Training Framework for Sparse CNN Accelerators

Yuanchen Qu, Yu Ma, Pingqiang Zhou

Abstract: Sparse convolution neural network (CNN) accelera-tors have shown to achieve high processing speed and low energy consumption by leveraging zero weights or activations, which can be further optimized by finely tuning the sparse activation maps in training process. In this paper, we propose a CNN training frame-work targeting at reducing energy consumption and processing cycles in sparse CNN accelerators. We first model accelerator's energy consumption and processing cycles as functions of layer-wise activation map sparsity. Then we leverage the model and propose a hybrid regularization approximation method to further sparsify activation maps in the training process. The results show that our proposed framework can reduce the energy consumption of Eyeriss by 31.33%, 20.6% and 26.6% respectively on MobileNet-V2, SqueezeNet and Inception-V3. In addition, the processing speed can be increased by <tex>$\boldsymbol{1.96\times, 1.4\times}$</tex> and <tex>$\boldsymbol{1.65}\times$</tex> respectively.