Decoupled Multi-task Learning with Cyclical Self-Regulation for Face Parsing.

The official repository of Decoupled Multi-task Learning with Cyclical Self-Regulation for Face Parsing. (CVPR 2022).

Installation

Our model is based on Pytorch 1.7.1 with Python 3.6.2.

pip install -r requirements.txt

Data

You can download original datasets:

• Helen : https://www.sifeiliu.net/face-parsing
• LaPa : https://github.com/JDAI-CV/lapa-dataset
• CelebAMask-HQ : https://github.com/switchablenorms/CelebAMask-HQ

and put them in ./dataset folder as below

dataset/
    images/
    labels/
    edges/
    train_list.txt
    test_list.txt
        each line: 'images/100032540_1.jpg labels/100032540_1.png'

Besides, we provide the edge genearation code in the generate_edge.py.

Usage

If you need imagenet pretrained resent-101, please download from baidu drive or Google drive, and put it into snapshot folder.

For dstributed(multi-gpu) training. Inplace-abn requires pytorch distributed data parallel.

GPU=4,5,6,7
Node=4
dataset=./datasets/CelebAMask-HQ/
snapshot=./work_dirs/
CUDA_VISIBLE_DEVICES="$GPU" python -m torch.distributed.launch --nproc_per_node="$Node"  --master_port=295002 train.py --data-dir "$dataset"  --random-mirror --random-scale \
--gpu "$GPU" --batch-size 7 --input-size 473,473 --snapshot-dir "$snapshot" --num-classes 19 --epochs 200 --schp-start 150

For testing pretrained models

python test.py --data-dir "$dataset" --out-dir "$out_dir" --restore-from "$snapshot" --gpu "$GPU" --batch-size 7 --input-size 473,473 --dataset test --num-classes 19

Reference

If you consider use our code, please cite our paper:

@inproceedings{Zheng2022DecoupledML,
  title={Decoupled Multi-task Learning with Cyclical Self-Regulation for Face Parsing},
  author={Qi Zheng and Jiankang Deng and Zheng Zhu and Ying Li and Stefanos Zafeiriou},
  booktitle={Computer Vision and Pattern Recognition},
  year={2022}
}