Despite the strong performance of transformers, quadratic computation complexity of self-attention presents challenges in applying them to vision tasks. Linear attention reduces this complexity from quadratic to linear, offering a strong computation-performance tradeoff. To further optimize this, automatic pruning is an effective method to find a structure that maximizes performance within a target resource through training without any heuristic approaches. However, directly applying it to multihead attention is not straightforward due to channel mismatch. In this article, we propose an automatic pruning method to deal with this problem. Different from existing methods that rely solely on training without any prior knowledge, we integrate channel similarity-based weights into the pruning indicator to preserve the more informative channels within each head. Then, we adjust the pruning indicator to enforce that channels are removed evenly across all heads, thereby avoiding any channel mismatch. We incorporate a reweight module to mitigate information loss due to channel removal and introduce an effective pruning indicator initialization for linear attention, based on the attention differences between the original structure and each channel. By applying our pruning method to the FLattenTransformer on ImageNet-1K, which incorporates original and linear attention mechanisms, we achieve a 30% reduction of FLOPs in a near lossless manner. It also has 1.96% of accuracy gain over the DeiT-B model while reducing FLOPs by 37%, and 1.05% accuracy increase over the Swin-B model with a 10% reduction in FLOPs as well. The proposed method outperforms previous state-of-the-art efficient models and the recent pruning methods.
Lee et al. (2025) studied this question.