GraNet 기반의 필터 프루닝을 적용한 경량 모델의 양자화 효과에 대한 연구

Abstract

As convolutional neural networks get deeper and wider, model compression is being widely used to reduce the amount of computation and memory usage. Pruning, which includes structured pruning and unstructured pruning, is one of the widely-adopted model compression methods. The structured pruning can reduce the size of the network model by model thinning, but it may suffer from worse accuracy degradation than the unstructured method. In this study, we claim that if quantization is used in conjunction with the structured pruning, the data size can be reduced without significantly sacrificing the model's performance. We propose a lightweight model on which both the GraNet structured pruning and an 8-bit weight quantization are applied. We evaluate the performance of both static and dynamic quantization to quantize the pruned model. The experiment was conducted to perform image classification tasks using the ResNet18 model with pruning and quantization on CIFAR-100 datasets. Compared to the original model, we reduced the weight size of the model by 84.25%, 88%, and 96.25% with constraints of 2.5%, 5%, and 10% accuracy degradation using GraNet filter pruning and 8-bit quantization.