Pruning Self-attentions into Convolutional Layers in Single Path

Vision Transformers (ViTs) have achieved impressive performance over various computer vision tasks. However, modeling global correlations with multi-head self-attention (MSA) layers leads to two widely recognized issues: the massive computational resource consumption and the lack of intrinsic inductive bias for modeling local visual patterns. One unified solution is to search whether to replace some MSA layers with convolution-like inductive biases that are computationally efficient via neural architecture search (NAS) based pruning methods. However, maintaining MSA and different candidate convolutional operations as separate trainable paths gives rise to expensive search cost and challenging optimization. Instead, we propose a novel weight-sharing scheme between MSA and convolutional operations and cast the search problem as finding which subset of parameters to use in each MSA layer. The weight-sharing scheme further allows us to devise an automatic Single-Path Vision Transformer pruning method (SPViT) to quickly prune the pre-trained ViTs into accurate and compact hybrid models with significantly reduced search cost, given target efficiency constraints. We conduct extensive experiments on two representative ViT models showing our method achieves a favorable accuracy-efficiency trade-off. Code is available at https://github.com/zhuang-group/SPViT.