Rotary Position Embedding for Vision Transformer

Byeongho Heo, Song Park, Dongyoon Han, Sangdoo Yun

NAVER AI LAB

Official PyTorch implementation of RoPE-ViT "Rotary Position Embedding for Vision Transformer" | arxiv, ECCV

Abstract

Rotary Position Embedding (RoPE) performs remarkably on language models, especially for length extrapolation of Transformers. However, the impacts of RoPE on computer vision domains have been underexplored, even though RoPE appears capable of enhancing Vision Transformer (ViT) performance in a way similar to the language domain. This study provides a comprehensive analysis of RoPE when applied to ViTs, utilizing practical implementations of RoPE for 2D vision data. The analysis reveals that RoPE demonstrates impressive extrapolation performance, i.e., maintaining precision while increasing image resolution at inference. It eventually leads to performance improvement for ImageNet-1k, COCO detection, and ADE-20k segmentation. We believe this study provides thorough guidelines to apply RoPE into ViT, promising improved backbone performance with minimal extra computational overhead.

Updates

• Oct 16, 2024: Pre-trained weights are uploaded to huggingface hub
• Apr 16, 2024: Pre-trained weights released
• Apr 15, 2024: Code released
• Mar 21, 2024: Arxiv paper was released

Getting Started

You can find RoPE implementations at each folder.

• deit/ : RoPE on DeiT-III training code "DeiT III: Revenge of the ViT" original repo
• swin/ : RoPE on Swin Transformer training code "Swin Transformer: Hierarchical Vision Transformer using Shifted Windows" original repo
• models/ : Only RoPE model files that used for DeiT and Swin.
• self-attn/ : Minimum implementation of RoPE in self-attention layer

Performances

DeiT-III

Swin Transformer

Pre-trained weights

• You can access the pre-trained weights on our huggingface hub
• pretrained=True for models in models/ will automatically load pre-trained weights from huggingface hub

  from models import vit_rope
  model = vit_rope.rope_mixed_deit_small_patch16_LS(pretrained=True)
  from models import swin_transformer_rope
  model = swin_transformer_rope.swin_rope_mixed_tiny_patch4_window7_224(pretrained=True)

• For inference with 256x256 images

  from models import vit_rope
  model = vit_rope.rope_mixed_deit_small_patch16_LS(pretrained=True, img_size=256)
  from models import swin_transformer_rope
  model = swin_transformer_rope.swin_rope_mixed_tiny_patch4_window7_224(pretrained=True, img_size=256)

DeiT-III (400 epochs)

Model Name	Top-1 (224)	Top-1 (384)	Weights
deit_small_patch16_LS	80.4	79.4	HF hub / Google drive
rope_axial_deit_small_patch16_LS	80.9	80.0	HF hub / Google drive
rope_mixed_deit_small_patch16_LS	80.9	81.8	HF hub / Google drive
rope_axial_ape_deit_small_patch16_LS	80.7	81.2	HF hub / Google drive
rope_mixed_ape_deit_small_patch16_LS	80.9	81.7	HF hub / Google drive
deit_base_patch16_LS	83.4	82.8	HF hub / Google drive
rope_axial_deit_base_patch16_LS	83.6	83.9	HF hub / Google drive
rope_mixed_deit_base_patch16_LS	83.8	84.4	HF hub / Google drive
rope_axial_ape_deit_base_patch16_LS	83.7	83.8	HF hub / Google drive
rope_mixed_ape_deit_base_patch16_LS	83.8	84.3	HF hub / Google drive
deit_large_patch16_LS	84.6	84.2	HF hub / Google drive
rope_axial_deit_large_patch16_LS	84.7	85.1	HF hub / Google drive
rope_mixed_deit_large_patch16_LS	84.8	85.6	HF hub / Google drive
rope_axial_ape_deit_large_patch16_LS	84.7	85.1	HF hub / Google drive
rope_mixed_ape_deit_large_patch16_LS	84.9	85.5	HF hub / Google drive

Swin Transformer (300 epochs)

Model Name	Top-1 (224)	Top-1 (384)	Weights
swin_tiny_patch4_window7_224	81.2	78.9
swin_rope_axial_tiny_patch4_window7_224	81.3	79.2	HF hub / Google drive
swin_rope_mixed_tiny_patch4_window7_224	81.4	79.5	HF hub / Google drive
swin_small_patch4_window7_224	82.9	81.0
swin_rope_axial_small_patch4_window7_224	83.1	80.9	HF hub / Google drive
swin_rope_mixed_small_patch4_window7_224	83.0	81.4	HF hub / Google drive
swin_base_patch4_window7_224	83.3	81.2
swin_rope_axial_base_patch4_window7_224	83.6	81.8	HF hub / Google drive
swin_rope_mixed_base_patch4_window7_224	83.7	82.1	HF hub / Google drive

How to cite

@inproceedings{heo2024ropevit,
    title={Rotary Position Embedding for Vision Transformer},
    author={Heo, Byeongho and Park, Song and Han, Dongyoon and Yun, Sangdoo},
    year={2024},
    booktitle={European Conference on Computer Vision (ECCV)},
}

License

This project is distributed under Apache-2.0,
except for the files below which originated from https://github.com/meta-llama/codellama.

• deit/models_v2_rope.py
• models/swin_transformer_rope.py
• models/vit_rope.py
• self-attn/rope_self_attn.py
• Swin-Transformer/models/swin_transformer_rope.py

RoPE-ViT
Copyright (c) 2024-present NAVER Cloud Corp.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.