A Knowledge- Distillation - Integrated Pruning Method for Vision Transformer

Vision transformers (ViTs) have made remarkable achievements in various computer vision applications such as image classification, object detection, and image segmentation. Since the self-attention mechanism introduced by itself can model the relationship between all pixels of the input image, the performance of the ViTs model is significantly improved compared to the traditional CNN network. However, their storage, runtime memory and computing requirements hinder their deployment on edge devices. This paper proposes a ViT pruning method with knowledge distillation, which can prune the ViT model and avoid the performance loss of the model after pruning. Based on the idea that knowledge distillation can make the student model improve the performance of the model by learning the unique knowledge of the teacher model, the convolution neural network (CNN) which has the unique ability of parameter sharing and local receptive field is used as a teacher model to guide the training of the ViT model and enable the ViT model to obtain the same ability. In addition, some important parts may be cut during pruning, resulting in irreversible loss of model performance. To solve this problem, this paper designs the importance score learning module to guide the pruning work, and determines that the pruning work removes the unimportant parts of the model. Finally, this paper compares the pruned model with other methods in terms of accuracy, Floating Point Operations(FLOPs) and model parameters on ImageNet-1k.