KKDnet

Kernel-mask Knowledge Distillation (KKD)

News

• (2022/05/20) KKDnet

Introduction

In this paper, we proposed a novel distillation method for efficient and accurate arbitrary-shaped text detection, termed Kernel-mask Knowledge Distillation (KKD), to improve the efficiency and accuracy of text detection.

The overall network architecture diagram will be uploaded after publication

Installation

First, clone the repository locally:

git clone https://github.com/giganticpower/KKDnet.git

required:

PyTorch 1.1.0+

torchvision 0.3.0+

pip install -r requirement.txt

compile codes of post-processing:

sh ./compile.sh

Dataset

Please refer to dataset/README.md for dataset preparation.

Training

first stage: we train the teacher network

CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py ${CONFIG_FILE}

For example:

CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py config/kkd/r50_ctw.py

second stage: knowledge distillation -> training the student network

CUDA_VISIBLE_DEVICES=0,1,2,3 python knowledge_distillation_train.py

NOTE: in knowledge_distillation_train.py, we should chose the dataloader in the code 162 - 184 lines

Testing

Evaluate

one checkpoint testing:

python test.py ${CONFIG_FILE} ${CHECKPOINT_FILE}

For example:

python test.py config/kkd/r18_ctw.py checkpoints/checkpoint.pth.tar

batch checkpoint testing:

python batch_eval.py ${CONFIG_FILE}

For example:

python batch_eval.py config/kkd/r18_ctw.py

note:

1.when batch checkpoint testing, you should change the checkpoint path in batch_eval.py

2.you should change the eval dataset in test.py 151-160 lines. (Here's a quick prediction link we've designed for convenience)

Citation

Please cite the related works in your publications if it helps your research:

KKDnet

Stay tuned

License

This project is developed and maintained by [Fuzhou University].