A Novel Deep Learning Model Compression Algorithm

In order to solve the problem of large model computing power consumption, this paper proposes a novel compression algorithm. Firstly, an interpretable weight allocation method for loss between student network (a with poor performance), teacher better performance) and real label. Then, different from previous simple pruning fine-tuning, performs knowledge distillation on pruned model, quantifies residual weights distilled model. The above operations can further reduce size calculation cost while maintaining accuracy. experimental results show that proposed in allocate relatively appropriate tags. On cifar-10 dataset, combining quantization memory resnet32 3726 KB 1842 KB, accuracy be kept at 93.28%, higher than original Compared similar algorithms, operation speed are greatly improved.