Awesome

STViT-R

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.

Object Detection and Instance Segmentation: See STViT-R-Object-Detection

Results on ImageNet

Usage

Installation

See get_started.md for a quick installation.

Training

python -m torch.distributed.launch --nproc_per_node <num-of-gpus-to-use> --master_port 12345  main.py \ 
--cfg <config-file> --data-path <imagenet-path> [--batch-size <batch-size-per-gpu> --output <output-directory> --tag <job-tag>]

For example, to train Swin Transformer with 8 GPU on a single node for 300 epochs, run:

STViT-R-Swin-S:

python -m torch.distributed.launch --nproc_per_node 8 --master_port 12345  main.py \
--cfg configs/swin_small_patch4_window7_224.yaml --data-path <imagenet-path> --batch-size 128 

STViT-R-Swin-B:

python -m torch.distributed.launch --nproc_per_node 8 --master_port 12345  main.py \
--cfg configs/swin_base_patch4_window7_224.yaml --data-path <imagenet-path> --batch-size 64 \
--accumulation-steps 2 [--use-checkpoint]

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}
}