ASR: Attention-alike Structural Re-parameterization
This work addresses a bottleneck in SRP techniques for deep learning practitioners, enabling broader industrial use by incorporating attention modules, though it is incremental as it builds on existing SRP methods.
The paper tackles the problem that structural re-parameterization (SRP) cannot handle attention modules due to their input-dependent nature, which limits SRP's practical applications; it proposes ASR, a method enabling SRP for attention mechanisms, and shows it improves performance on standard benchmarks without extra model crafting.
The structural re-parameterization (SRP) technique is a novel deep learning technique that achieves interconversion between different network architectures through equivalent parameter transformations. This technique enables the mitigation of the extra costs for performance improvement during training, such as parameter size and inference time, through these transformations during inference, and therefore SRP has great potential for industrial and practical applications. The existing SRP methods have successfully considered many commonly used architectures, such as normalizations, pooling methods, and multi-branch convolution. However, the widely used attention modules which drastically slow inference speed cannot be directly implemented by SRP due to these modules usually act on the backbone network in a multiplicative manner and the modules' output is input-dependent during inference, which limits the application scenarios of SRP. In this paper, we conduct extensive experiments from a statistical perspective and discover an interesting phenomenon Stripe Observation, which reveals that channel attention values quickly approach some constant vectors during training. This observation inspires us to propose a simple-yet-effective attention-alike structural re-parameterization (ASR) that allows us to achieve SRP for a given network while enjoying the effectiveness of the attention mechanism. Extensive experiments conducted on several standard benchmarks demonstrate the effectiveness of ASR in generally improving the performance of existing backbone networks, attention modules, and SRP methods without any elaborated model crafting. We also analyze the limitations and provide experimental and theoretical evidence for the strong robustness of the proposed ASR.