CSwin

CSWin Transformer: A General Vision Transformer Backbone with Cross-Shaped Windows, arxiv

PaddlePaddle training/validation code and pretrained models for CSWin Transformer.

The official pytorch implementation is here.

This implementation is developed by PaddleViT.

CSWin Transformer Model Overview

Update

• Update (2022-03-16): Code is refactored.
• Update (2021-09-27): Model FLOPs and # params are uploaded.
• Update (2021-08-11): Code is released and ported weights are uploaded.

Models Zoo

Model	Acc@1	Acc@5	#Params	FLOPs	Image Size	Crop_pct	Interpolation	Link
cswin_tiny_224	82.81	96.30	22.3M	4.2G	224	0.9	bicubic	google/baidu
cswin_small_224	83.60	96.58	34.6M	6.5G	224	0.9	bicubic	google/baidu
cswin_base_224	84.23	96.91	77.4M	14.6G	224	0.9	bicubic	google/baidu
cswin_base_384	85.51	97.48	77.4M	43.1G	384	1.0	bicubic	google/baidu
cswin_large_224	86.52	97.99	173.3M	32.5G	224	0.9	bicubic	google/baidu
cswin_large_384	87.49	98.35	173.3M	96.1G	384	1.0	bicubic	google/baidu

*The results are evaluated on ImageNet2012 validation set.

For finetuning using 22k model, the ported weight file can be downloaded from:

• cswin_large_22k_224 google/baidu

Data Preparation

ImageNet2012 dataset is used in the following file structure:

│imagenet/
├──train_list.txt
├──val_list.txt
├──train/
│  ├── n01440764
│  │   ├── n01440764_10026.JPEG
│  │   ├── n01440764_10027.JPEG
│  │   ├── ......
│  ├── ......
├──val/
│  ├── n01440764
│  │   ├── ILSVRC2012_val_00000293.JPEG
│  │   ├── ILSVRC2012_val_00002138.JPEG
│  │   ├── ......
│  ├── ......

• train_list.txt: list of relative paths and labels of training images. You can download it from: google/baidu
• val_list.txt: list of relative paths and labels of validation images. You can download it from: google/baidu

Usage

To use the model with pretrained weights, download the .pdparam weight file and change related file paths in the following python scripts. The model config files are located in ./configs/.

For example, assume weight file is downloaded in ./cswin_tiny_224.pdparams, to use the cswin_tiny_224 model in python:

from config import get_config
from cswin import build_cswin as build_model
# config files in ./configs/
config = get_config('./configs/cswin_tiny_224.yaml')
# build model
model = build_model(config)
# load pretrained weights
model_state_dict = paddle.load('./cswin_tiny_224.pdparams')
model.set_state_dict(model_state_dict)

Evaluation

To evaluate CSwin model performance on ImageNet2012, run the following script using command line:

sh run_eval_multi.sh

or

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
python main_multi_gpu.py \
-cfg='./configs/cswin_tiny_224.yaml' \
-dataset='imagenet2012' \
-batch_size=256 \
-data_path='/dataset/imagenet' \
-eval \
-pretrained='./cswin_tiny_224.pdparams' \
-amp

Note: if you have only 1 GPU, change device number to CUDA_VISIBLE_DEVICES=0 would run the evaluation on single GPU.

Training

To train the CSwin model on ImageNet2012, run the following script using command line:

sh run_train_multi.sh

or

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
python main_multi_gpu.py \
-cfg='./configs/cswin_tiny_224.yaml' \
-dataset='imagenet2012' \
-batch_size=256 \
-data_path='/dataset/imagenet' \
-amp

Note: it is highly recommanded to run the training using multiple GPUs / multi-node GPUs.

Finetuning

To finetune the Swin model on ImageNet2012, run the following script using command line:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
python main_multi_gpu.py \
-cfg='./configs/cswin_base_384.yaml' \
-dataset='imagenet2012' \
-batch_size=16 \
-data_path='/dataset/imagenet' \
-pretrained='./cswin_base_224.pdparams' \
-amp

Note: use -pretrained argument to set the pretrained model path, you may also need to modify the hyperparams defined in config file.

Reference

@article{dong2021cswin,
  title={CSWin Transformer: A General Vision Transformer Backbone with Cross-Shaped Windows},
  author={Dong, Xiaoyi and Bao, Jianmin and Chen, Dongdong and Zhang, Weiming and Yu, Nenghai and Yuan, Lu and Chen, Dong and Guo, Baining},
  journal={arXiv preprint arXiv:2107.00652},
  year={2021}
}