Visformer: The Vision-friendly Transformer

TL;DR

Visformer is a new architecture that combines the strengths of Transformers and convolutional models, improving image classification accuracy especially with limited data and lower complexity.

Contribution

The paper introduces Visformer, a novel vision-friendly Transformer architecture that outperforms existing models in accuracy and robustness, especially with smaller datasets.

Findings

Visformer achieves higher ImageNet accuracy than Transformer and convolution models.

Visformer performs better with smaller training sets and lower model complexity.

The transition study reveals insights into the overfitting issues of Transformer models.

Abstract

The past year has witnessed the rapid development of applying the Transformer module to vision problems. While some researchers have demonstrated that Transformer-based models enjoy a favorable ability of fitting data, there are still growing number of evidences showing that these models suffer over-fitting especially when the training data is limited. This paper offers an empirical study by performing step-by-step operations to gradually transit a Transformer-based model to a convolution-based model. The results we obtain during the transition process deliver useful messages for improving visual recognition. Based on these observations, we propose a new architecture named Visformer, which is abbreviated from the `Vision-friendly Transformer'. With the same computational complexity, Visformer outperforms both the Transformer-based and convolution-based models in terms of ImageNet classification accuracy, and the advantage becomes more significant when the model complexity is lower or the training set is smaller. The code is available at https://github.com/danczs/Visformer.