- python3.7
- pytorch1.5.0 + torchvision0.6.0
- CUDA
- opencv-python
- numpy
- tensorboardX
- tqdm

# 1. Create a conda virtual environment.
conda create -n ffwm python=3.7 anaconda
source activate ffwm

# 2. Install the pytorch 
conda install pytorch==1.5.0 torchvision==0.6.0 cudatoolkit=9.2 -c pytorch

# 3. Install dependency
pip install -r requirements.txt

# 4. Build pytorch Custom CUDA Extensions, we have tested it on pytorch1.5.0+cu92
bash setup.sh

dataset
    ├── multipie
    │       ├── train
    │       │     ├── images
    │       │     ├── masks
    │       │     └── landmarks.npy
    │       └── test
    │             ├── images
    │             ├── gallery_list.npy (optional)
    │             └── visual_list.npy (optional)
    └── lfw
         ├── images
         └── pairs.txt

./checkpoints
      ├── ffwm
      │       ├── latest_net_flowNetF.pth
      │       └── latest_net_netG.pth
      ├── lightCNN_10_checkpoint.pth (pretrained)
      └── LightCNN_29Layers_checkpoint.pth (original)

python test_ffwm.py \
            --dataroot path/to/dataset \
            --lightcnn path/to/pretrained lightcnn \
            --preload 

python test_ffwm.py \
            --datamode lfw \
            --dataroot path/to/dataset \
            --lightcnn path/to/pretrained lightcnn \
            --preload 

cd lightcnn
python finetune.py \
            --save_path ../checkpoints/ \
            --dataroot path/to/dataset/multipie \
            --model_path path/to/original lightcnn \
            --preload

python train_flow.py \
            --model flownet \
            --dataroot path/to/dataset \
            --aug \
            --preload \
            --name flownetf \
            --batch_size 6

python train_flow.py \
            --model flownet \
            --reverse \
            --dataroot path/to/dataset \
            --aug \
            --preload \
            --name flownetb \
            --batch_size 6

python train_ffwm.py \
                --name ffwm  \
                --preload \
                --dataroot path/to/dataset \
                --lightcnn path/to/pretrained lightcnn 

@InProceedings{wei2020ffwm,
  author = {Wei, Yuxiang and Liu, Ming and Wang, Haolin and Zhu, Ruifeng and Hu, Guosheng and Zuo, Wangmeng},
  title = {Learning Flow-based Feature Warping For Face Frontalization with Illumination Inconsistent Supervision},
  booktitle = {Proceedings of the European Conference on Computer Vision},
  year = {2020}
}