This repo is the official implementation of "Refiner: Refining Self-attention for Vision Transformers". The repo is build on top of timm and include the relabbeling trick included in TokenLabelling.

Refined Vision Transformer is initially described in arxiv, which observes vision transformers require much more datafor model pre-training. Most of recent works thus are dedicated to designing morecomplex architectures or training methods to address the data-efficiency issue ofViTs. However, few of them explore improving the self-attention mechanism, akey factor distinguishing ViTs from CNNs. Different from existing works, weintroduce a conceptually simple scheme, calledrefiner, to directly refine the self-attention maps of ViTs. Specifically, refiner exploresattention expansionthatprojects the multi-head attention maps to a higher-dimensional space to promotetheir diversity. Further, refiner applies convolutions to augment local patternsof the attention maps, which we show is equivalent to adistributed local atten-tion—features are aggregated locally with learnable kernels and then globallyaggregated with self-attention. Extensive experiments demonstrate that refinerworks surprisingly well. Significantly, it enables ViTs to achieve 86% top-1 classifi-cation accuracy on ImageNet with only 81M parameters.

Please run git clone with --recursive to clone timm as submodule and install it with cd pytorch-image-models && pip install -e ./

torch>=1.4.0 torchvision>=0.5.0 pyyaml numpy timm==0.4.5

A summary of the results are shown below for quick reference. Details can be found in the paper.

Train the Refiner-ViT-S from scratch:

bash run.sh scripts/refiner_s.yaml 

To use the re-labbeling tricks for improving the accuracy, download the relabel_data based on NFNet. This is provided in TokenLabelling repo. Then, copy the relabbeling data to the data folder.