Edge-aware Graph Representation Learning and Reasoning for Face Parsing

The official repository of Edge-aware Graph Representation Learning and Reasoning for Face Parsing (ECCV 2020).

Installation

Our model is based on Pytorch 1.4.0 with Python 3.6.8. Also, we use In-Place Activated BatchNorm. First, you need to clone and compile inplace_abn.

git clone https://github.com/mapillary/inplace_abn.git
cd inplace_abn
python setup.py install
cd scripts
pip install -r requirements.txt

Data

You can download the Helen dataset on https://www.sifeiliu.net/face-parsing and imagenet pretrained resent-101 from baidu drive or Google drive, and put it into snapshot folder. We do not provide the registration code for the moment, and you need to organize input data as follows:

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

We support single-gpu and multi-gpu training. Inplace-abn requires pytorch distributed data parallel.

Single gpu training

python train.py --data-dir ./dataset/Helen/ --random-mirror --random-scale --gpu 0 --learning-rate 1e-3 --weight-decay 5e-4 --batch-size 7 --input-size 473,473 --snapshot-dir ./snapshots/ --dataset train --num-classes 11 --epochs 200

Distributed(multi-gpu) training

CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 train.py --data-dir ./dataset/Helen/ --random-mirror --random-scale --gpu 0,1,2,3 --learning-rate 1e-3 --weight-decay 5e-4 --batch-size 7 --input-size 473,473 --snapshot-dir ./snapshots/ --dataset train --num-classes 11 --epochs 99

Validation

python evaluate.py --data-dir ./dataset/Helen/ --restore-from ./snapshots/helen/best.pth --gpu 0 --batch-size 7 --input-size 473,473 --dataset test --num-classes 11

Reference

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

@article{te2020edge,
  title={Edge-aware Graph Representation Learning and Reasoning for Face Parsing},
  author={Te, Gusi and Liu, Yinglu and Hu, Wei and Shi, Hailin and Mei, Tao},
  journal={arXiv preprint arXiv:2007.11240},
  year={2020}
}