CSWin-Transformer
This repo is the official implementation of "CSWin Transformer: A General Vision Transformer Backbone with Cross-Shaped Windows". The code and models for downstream tasks are coming soon.
Introduction
CSWin Transformer (the name CSWin
stands for Cross-Shaped Window) is introduced in arxiv, which is a new general-purpose backbone for computer vision. It is a hierarchical Transformer and replaces the traditional full attention with our newly proposed cross-shaped window self-attention. The cross-shaped window self-attention mechanism computes self-attention in the horizontal and vertical stripes in parallel that from a cross-shaped window, with each stripe obtained by splitting the input feature into stripes of equal width. With CSWin, we could realize global attention with a limited computation cost.
CSWin Transformer achieves strong performance on ImageNet classification (87.5 on val with only 97G flops) and ADE20K semantic segmentation (55.7 mIoU
on val), surpassing previous models by a large margin.
Main Results on ImageNet
model | pretrain | resolution | acc@1 | #params | FLOPs | 22K model | 1K model |
---|---|---|---|---|---|---|---|
CSWin-T | ImageNet-1K | 224x224 | 82.8 | 23M | 4.3G | - | model |
CSWin-S | ImageNet-1k | 224x224 | 83.6 | 35M | 6.9G | - | model |
CSWin-B | ImageNet-1k | 224x224 | 84.2 | 78M | 15.0G | - | model |
CSWin-B | ImageNet-1k | 384x384 | 85.5 | 78M | 47.0G | - | model |
CSWin-L | ImageNet-22k | 224x224 | 86.5 | 173M | 31.5G | model | model |
CSWin-L | ImageNet-22k | 384x384 | 87.5 | 173M | 96.8G | - | model |
Main Results on Downstream Tasks
COCO Object Detection
backbone | Method | pretrain | lr Schd | box mAP | mask mAP | #params | FLOPS |
---|---|---|---|---|---|---|---|
CSwin-T | Mask R-CNN | ImageNet-1K | 3x | 49.0 | 43.6 | 42M | 279G |
CSwin-S | Mask R-CNN | ImageNet-1K | 3x | 50.0 | 44.5 | 54M | 342G |
CSwin-B | Mask R-CNN | ImageNet-1K | 3x | 50.8 | 44.9 | 97M | 526G |
CSwin-T | Cascade Mask R-CNN | ImageNet-1K | 3x | 52.5 | 45.3 | 80M | 757G |
CSwin-S | Cascade Mask R-CNN | ImageNet-1K | 3x | 53.7 | 46.4 | 92M | 820G |
CSwin-B | Cascade Mask R-CNN | ImageNet-1K | 3x | 53.9 | 46.4 | 135M | 1004G |
ADE20K Semantic Segmentation (val)
Backbone | Method | pretrain | Crop Size | Lr Schd | mIoU | mIoU (ms+flip) | #params | FLOPs |
---|---|---|---|---|---|---|---|---|
CSwin-T | Semantic FPN | ImageNet-1K | 512x512 | 80K | 48.2 | - | 26M | 202G |
CSwin-S | Semantic FPN | ImageNet-1K | 512x512 | 80K | 49.2 | - | 39M | 271G |
CSwin-B | Semantic FPN | ImageNet-1K | 512x512 | 80K | 49.9 | - | 81M | 464G |
CSwin-T | UPerNet | ImageNet-1K | 512x512 | 160K | 49.3 | 50.4 | 60M | 959G |
CSwin-S | UperNet | ImageNet-1K | 512x512 | 160K | 50.0 | 50.8 | 65M | 1027G |
CSwin-B | UperNet | ImageNet-1K | 512x512 | 160K | 50.8 | 51.7 | 109M | 1222G |
CSwin-B | UPerNet | ImageNet-22K | 640x640 | 160K | 51.8 | 52.6 | 109M | 1941G |
CSwin-L | UperNet | ImageNet-22K | 640x640 | 160K | 53.4 | 55.7 | 208M | 2745G |
Requirements
timm==0.3.4, pytorch>=1.4, opencv, ... , run:
bash install_req.sh
Apex for mixed precision training is used for finetuning. To install apex, run:
git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
Data prepare: ImageNet with the following folder structure, you can extract imagenet by this script.
│imagenet/
├──train/
│ ├── n01440764
│ │ ├── n01440764_10026.JPEG
│ │ ├── n01440764_10027.JPEG
│ │ ├── ......
│ ├── ......
├──val/
│ ├── n01440764
│ │ ├── ILSVRC2012_val_00000293.JPEG
│ │ ├── ILSVRC2012_val_00002138.JPEG
│ │ ├── ......
│ ├── ......
Train
Train the three lite variants: CSWin-Tiny, CSWin-Small and CSWin-Base:
bash train.sh 8 --data <data path> --model CSWin_64_12211_tiny_224 -b 256 --lr 2e-3 --weight-decay .05 --amp --img-size 224 --warmup-epochs 20 --model-ema-decay 0.99984 --drop-path 0.2
bash train.sh 8 --data <data path> --model CSWin_64_24322_small_224 -b 256 --lr 2e-3 --weight-decay .05 --amp --img-size 224 --warmup-epochs 20 --model-ema-decay 0.99984 --drop-path 0.4
bash train.sh 8 --data <data path> --model CSWin_96_24322_base_224 -b 128 --lr 1e-3 --weight-decay .1 --amp --img-size 224 --warmup-epochs 20 --model-ema-decay 0.99992 --drop-path 0.5
If you want to train our CSWin on images with 384x384 resolution, please use '--img-size 384'.
If the GPU memory is not enough, please use '-b 128 --lr 1e-3 --model-ema-decay 0.99992' or use checkpoint '--use-chk'.
Finetune
Finetune CSWin-Base with 384x384 resolution:
bash finetune.sh 8 --data <data path> --model CSWin_96_24322_base_384 -b 32 --lr 5e-6 --min-lr 5e-7 --weight-decay 1e-8 --amp --img-size 384 --warmup-epochs 0 --model-ema-decay 0.9998 --finetune <pretrained 224 model> --epochs 20 --mixup 0.1 --cooldown-epochs 10 --drop-path 0.7 --ema-finetune --lr-scale 1 --cutmix 0.1
Finetune ImageNet-22K pretrained CSWin-Large with 224x224 resolution:
bash finetune.sh 8 --data <data path> --model CSWin_144_24322_large_224 -b 64 --lr 2.5e-4 --min-lr 5e-7 --weight-decay 1e-8 --amp --img-size 224 --warmup-epochs 0 --model-ema-decay 0.9996 --finetune <22k-pretrained model> --epochs 30 --mixup 0.01 --cooldown-epochs 10 --interpolation bicubic --lr-scale 0.05 --drop-path 0.2 --cutmix 0.3 --use-chk --fine-22k --ema-finetune
If the GPU memory is not enough, please use checkpoint '--use-chk'.
Cite CSWin Transformer
@misc{dong2021cswin,
title={CSWin Transformer: A General Vision Transformer Backbone with Cross-Shaped Windows},
author={Xiaoyi Dong and Jianmin Bao and Dongdong Chen and Weiming Zhang and Nenghai Yu and Lu Yuan and Dong Chen and Baining Guo},
year={2021},
eprint={2107.00652},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
Acknowledgement
This repository is built using the timm library and the DeiT repository.
License
This project is licensed under the license found in the LICENSE file in the root directory of this source tree.
Microsoft Open Source Code of Conduct
Contact Information
For help or issues using CSWin Transformer, please submit a GitHub issue.
For other communications related to CSWin Transformer, please contact Jianmin Bao ([email protected]
), Dong Chen ([email protected]
).