HFPQ: deep neural network compression by hardware-friendly pruning-quantization

This paper presents a hardware-friendly compression method for deep neural networks. This method effectively combines layered channel pruning with quantization by a power exponential of 2. While keeping a small decrease in the accuracy of the network model, the computational resources for neural networks to be deployed on the hardware are greatly reduced. These computing resources for hardware resolution include memory, multiple accumulation cells (MACs), and many logic gates for neural networks. Layered channel pruning groups the different layers by decreasing the model accuracy of the pruned network. After pruning each layer in a specific order, the network is retrained. The pruning method in this paper sets a parameter, that can be adjusted to meet different pruning rates in practical applications. The quantization method converts high-precision weights to low-precision weights. The latter are all composed of 0 and powers of 2. In the same way, another parameter is set to control the quantized bit width, which can also be adjusted to meet different quantization precisions. The hardware-friendly pruning quantization (HFPQ) method proposed in this paper trains the network after pruning and then quantizes the weights. The experimental results show that the HFPQ method compresses VGGNet, ResNet and GoogLeNet by 30+ times while reducing the number of FLOPs by more than 85%.