SP-ViT

This is the official implementation of our BMVC 2022 paper SP-ViT: Learning 2D Spatial Priors for Vision Transformers.

Comparison between our proposed SP-ViT and state-of-the-art vision transformers. Note that we exclude models pretrained on extra data or larger resolution than 224 × 224 for a fair comparison.

Our implementation is based on the TokenLabeling and pytorch-image-models.

SP-ViT Models(Using TokenLabeling as baseline)

Model	layer	dim	Image resolution	Param	Top 1	Download
SP-ViT-S	16	384	224	26M	83.9	link
SP-ViT-S	16	384	384	26M	85.1	link
SP-ViT-M	20	512	224	56M	84.9	link
SP-ViT-M	20	512	384	56M	86.0	link
SP-ViT-L	24	768	224	150M	85.5	link
SP-ViT-L	24	768	384	150M	86.3	link

Requirements

pyyaml scipy timm==0.4.5

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
│  │   ├── ......
│  ├── ......

Validation

Replace DATA_DIR with your imagenet validation set path and MODEL_DIR with the checkpoint path

CUDA_VISIBLE_DEVICES=0 bash eval.sh /path/to/imagenet/val /path/to/checkpoint

Label data

Please go to TokenLabeling for downloading the token label data.

Training

Train SP-ViT:

bash distributed_train.sh

Fine-tuning

To Fine-tune the pre-trained SP-ViT on images with 384x384 resolution:

bash distributed_fine_tune.sh

Segmentation

Please refer to TokenLabeling for more details:

Backbone	Method	Crop size	Lr Schd	mIoU	mIoU(ms)	Pixel Acc.	Param
LV-ViT-S	UperNet	512x512	160k	47.9	48.6	83.1	44M
SP-ViT-S	UperNet	512x512	160k	49.0	49.8	83.4	44M

Visualization

We apply the visualization method [Transformer-Explainability] (https://github.com/hila-chefer/Transformer-Explainability) to visualize the parts of the image that led to a certain classification for DeiT-S and our SP-ViT-S (w/o TokenLabeling). The parts of the image that used by the network to make the decision are highlighted in red.

Label generation

To generate token label data for training:

python3 generate_label.py /path/to/imagenet/train /path/to/save/label_top5_train_nfnet --model dm_nfnet_f6 --pretrained --img-size 576 -b 32 --crop-pct 1.0

Reference

If you use this repo or find it useful, please consider citing:

@inproceedings{BMVC2022,
  author={Zhou, Yuxuan and Xiang, Wangmeng and Li, Chao and Wang, Biao and Wei, Xihan and Zhang, Lei and Keuper, Margret and Hua, Xiansheng},
 booktitle = {The 33rd British Machine Vision Conference},
 title = {SP-ViT: Learning 2D Spatial Priors for Vision Transformers},
 url = {https://bmvc2022.mpi-inf.mpg.de/0564.pdf},
 year = {2022}
}

Related projects

MMSegmentation, Transformer Explainability.