Transformer Knowledge Distillation for Efficient Semantic Segmentation [arxiv]

Structure: TransKD

Introduction

We propose the structural framework, TransKD, to distill the knowledge from feature maps and patch embeddings of vision transformers. TransKD enables the non-pretrained vision transformers perform on-par with the pretrained ones.

(a)-(c) Knowledge distillation in computer vision is split into three categories: response-based knowledge distillation, feature-based knowledge distillation, and relation-based knowledge distillation. (d) TransKD extracts the relation-based knowledge of feature maps and transformer-specific patch embedding knowledge at each stage.

Requirements

Environment: create a conda environment and activate it

conda create -n TransKD python=3.6
conda activate TransKD

Additional python pachages: poly scheduler and

pytorch == 1.7.1+cu92
torchvision == 0.8.2+cu92
mmsegmentation == 0.15.0
mmcv-full == 1.3.10
numpy
visdom

Or simply:

pip install -r requirements.txt

Datasets:

Cityscapes: download gtFine_trainvaltest.zip and leftImg8bit_trainvaltest.zip from cityscapes official website, then prepare the 19-class label with the createTrainIdLabelImgs.py from cityscapesscripts.
ACDC： download gt_trainval.zip and rgb_anon_trainvaltest.zip from ACDC official website.

Model Zoo

Download the following weights to the folder \outputs. Cityscapes:

Network	#Params(M)	GFLOPs	mIoU(%)	weight
Teacher(B2)	27.36	113.84	76.49	Google Drive
Student(B0)	3.72	13.67	55.86	Google Drive
+TransKD-Base	4.56	16.47	68.58	Google Drive
+TransKD-GL	5.22	16.80	68.87	Google Drive
+TransKD-EA	5.53	17.84	68.98	Google Drive

PVTv2 teacher

ACDC:

Network	mIoU(%)	weight
Teacher(B2)	69.34	Google Drive
Student(B0)	46.26	Google Drive
+TransKD-Base	58.56	Google Drive
+TransKD-GL	58.13	Google Drive
+TransKD-EA	59.09	Google Drive

Training

Download pretrained weights (SegFormer and PVTv2) to the folder \train\ckpt_pretained\.

cd train
CUDA_VISIBLE_DEVICES=0 python TransKDBase.py --dataset cityscapes --dataset /path/to/cityscapes #--dataset ACDC --dataset /path/to/ACDC
CUDA_VISIBLE_DEVICES=0 python TransKD_GLMixer.py --dataset cityscapes --dataset /path/to/cityscapes #--dataset ACDC --dataset /path/to/ACDC
CUDA_VISIBLE_DEVICES=0 python TransKD_EA.py --dataset cityscapes --dataset /path/to/cityscapes #--dataset ACDC --dataset /path/to/ACDC

Evaluation

Download trained weights (Google Drive) to the folder \outputs.

cd eval
CUDA_VISIBLE_DEVICES=0 python eval_cityscapes_iou.py --distillation-type TransKDBase
CUDA_VISIBLE_DEVICES=0 python eval_ACDC_iou.py --distillation-type TransKDBase
# distillation-type can be choices=['teacher','student','TransKDBase','TransKD_GL','TransKD_EA']

Inference

Download trained weights (Google Drive) to the folder \outputs.

cd eval
CUDA_VISIBLE_DEVICES=0 python eval_cityscapes_color.py --distillation-type TransKDBase
CUDA_VISIBLE_DEVICES=0 python eval_ACDC_color.py --distillation-type TransKDBase
# distillation-type can be choices=['teacher','student','TransKDBase','TransKD_GL','TransKD_EA']

Acknowledgement

Our framework is build upon Knowledge Review and Erfnet. Thanks to their superior work!

Publication

If you find this repo useful, please consider referencing the following paper [PDF]:

@article{liu2022transkd,
  title={TransKD: Transformer Knowledge Distillation for Efficient Semantic Segmentation},
  author={Liu, Ruiping and Yang, Kailun and Roitberg, Alina and Zhang, Jiaming and Peng, Kunyu and Liu, Huayao and Stiefelhagen, Rainer},
  journal={arXiv preprint arXiv:2202.13393},
  year={2022}
}

RuipingL / TransKD