Multi-Dimensional Vision Transformer Compression via Dependency Guided Gaussian Process Search
Vision transformers (ViT) have recently attracted considerable attentions, but the huge computational cost remains an issue for practical deployment. Previous ViT pruning methods tend to prune the model along one dimension solely, which may suffer from excessive reduction and lead to sub-optimal model quality. In contrast, we advocate a multi-dimensional ViT compression paradigm, and propose to harness the redundancy reduction from attention head, neuron and sequence dimensions jointly. Firstly, we propose a statistical dependence based pruning criterion that is generalizable to different dimensions for identifying the deleterious components. Moreover, we cast the multidimensional ViT compression as an optimization problem, objective of which is to learn an optimal pruning policy across the three dimensions while maximizing the compressed model's accuracy under a computational budget. The problem is solved by an adapted Gaussian process search with expected improvement. Experimental results show that our method effectively reduces the computational cost of various ViT models. For example, our method reduces 40% FLOPs without top-1 accuracy loss for DeiT and T2T-ViT models on the ImageNet dataset, outperforming previous state-of-the-art ViT pruning methods.