Scaled ReLU Matters for Training Vision Transformers

TL;DR

This paper demonstrates that using scaled ReLU in the convolutional stem of vision transformers significantly stabilizes training and enhances performance, revealing that proper architectural choices are crucial for ViT training success.

Contribution

It introduces the importance of scaled ReLU in the conv-stem for stable training and improved performance of vision transformers, supported by theoretical and empirical analysis.

Findings

Scaled ReLU in conv-stem stabilizes training.

Enhanced diversity of patch tokens boosts performance.

Previous ViTs are undertrained, indicating potential for improvement.

Abstract

Vision transformers (ViTs) have been an alternative design paradigm to convolutional neural networks (CNNs). However, the training of ViTs is much harder than CNNs, as it is sensitive to the training parameters, such as learning rate, optimizer and warmup epoch. The reasons for training difficulty are empirically analysed in ~\cite{xiao2021early}, and the authors conjecture that the issue lies with the \textit{patchify-stem} of ViT models and propose that early convolutions help transformers see better. In this paper, we further investigate this problem and extend the above conclusion: only early convolutions do not help for stable training, but the scaled ReLU operation in the \textit{convolutional stem} (\textit{conv-stem}) matters. We verify, both theoretically and empirically, that scaled ReLU in \textit{conv-stem} not only improves training stabilization, but also increases the diversity of patch tokens, thus boosting peak performance with a large margin via adding few parameters and flops. In addition, extensive experiments are conducted to demonstrate that previous ViTs are far from being well trained, further showing that ViTs have great potential to be a better substitute of CNNs.