STViT-R for Detection

This is an official implementation for Making Vision Transformers Efficient from A Token Sparsification View. It is based on Swin Transformer.

Notes: We will further clean the code and release the checkpoints in the future.

Results on COCO

Model	$AP^b$	$AP^b_{50}$	$AP^b_{75}$	$AP^b_s$	$AP^m$	$AP^m_{50}$	$AP^m_{75}$	$AP^m_{s}$	log
STViT-R-Swin-S	51.8	70.6	56.1	36.7	44.7	67.8	48.6	29.0	Link
STViT-R-Swin-B	52.2	70.8	56.8	36.5	45.2	68.3	49.1	29.5	Link

Usage

Installation

Please refer to MMDetection for installation and dataset preparation.

We use apex for mixed precision training by default. 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" ./

Training

To train a Cascade Mask R-CNN model with a STVIT-R-Swin-S backbone and 8 gpus, run:

tools/dist_train.sh configs/swin/cascade_mask_rcnn_swin_small_patch4_window7_mstrain_480-800_giou_4conv1f_adamw_3x_coco.py 8 --cfg-options model.pretrained=<PRETRAIN_MODEL> 

To train a Cascade Mask R-CNN model with a STVIT-R-Swin-B backbone and 8 gpus, run:

tools/dist_train.sh configs/swin/cascade_mask_rcnn_swin_base_patch4_window7_mstrain_480-800_giou_4conv1f_adamw_3x_coco.py 8 --cfg-options model.pretrained=<PRETRAIN_MODEL> 

Citing STViT-R

@inproceedings{chang2023making,
  title={Making Vision Transformers Efficient from A Token Sparsification View},
  author={Chang, Shuning and Wang, Pichao and Lin, Ming and Wang, Fan and Zhang, David Junhao and Jin, Rong and Shou, Mike Zheng},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={6195--6205},
  year={2023}
}