# PyramidTNT: Improved Transformer-in-Transformer Baselines with Pyramid Architecture

**Authors:** Kai Han, Jianyuan Guo, Yehui Tang, Yunhe Wang  
**arXiv:** [2201.00978](https://arxiv.org/abs/2201.00978)  
**Date:** 2022-01-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2201.00978v1/

## Key points

- Transformer networks have made significant strides in computer vision tasks.
- The Transformer-in-Transformer (TNT) architecture is notable for its ability to extract both local and global representations using inner and outer transformers.
- Researchers including Kai Han, Jianyuan Guo, Yehui Tang, and Yunhe Wang have introduced new TNT baselines with a pyramid architecture and a convolutional stem.
- The "PyramidTNT" model is an improvement over the original TNT, establishing hierarchical representations and outperforming leading vision transformers like Swin Transformer.
- The code for PyramidTNT will be available at https://github.com/huawei-noah/CV-Backbones/tree/master/tnt_pytorch.
- This work extends the "Transformer in Transformer" concept and highlights the ongoing evolution of transformer-based approaches in computer vision.

## Summary

In the realm of computer vision tasks, Transformer networks have made significant strides. The Transformer-in-Transformer (TNT) architecture stands out for its ability to extract both local and global representations using inner and outer transformers. Building upon this foundation, a team of researchers including Kai Han, Jianyuan Guo, Yehui Tang, and Yunhe Wang have introduced new TNT baselines featuring two innovative designs: a pyramid architecture and a convolutional stem. The resulting "PyramidTNT" model represents a substantial improvement over the original TNT by establishing hierarchical representations. Notably, PyramidTNT has demonstrated superior performance compared to leading vision transformers like Swin Transformer. The researchers anticipate that this enhanced baseline will prove invaluable for advancing research and practical applications in the field of vision transformers. For those interested in exploring further, the code for PyramidTNT is set to be made available at https://github.com/huawei-noah/CV-Backbones/tree/master/tnt_pytorch. This work serves as an extension of the "Transformer in Transformer" concept and underscores the ongoing evolution of transformer-based approaches in computer vision.