FaceForencics Detection

Use FaceForencics++ dataset to train model, get total 0.62 in benchmark now.

Model

Use Resnext50 to train NeuralTextures, Deepfakes and Original data, use some data augmentation process.

Train

Environment

# install pytorch with corresponding cuda version
pip install torch

# install facenet-pytorch
pip install facenet-pytorch

Get Train Data

Get FaceForencics++ dataset, put them to dataset folder. My folder structure is like following.

--- dataset
    --- faceforensics_benchmark_images
    --- mtcnn
        --- deepfakes_faces_c23
        --- deepfakes_faces_c40
        --- face2face_faces_c23
        --- face2face_faces_c40
        --- faceswap_faces_c23
        --- faceswap_faces_c40
        --- faceswap_faces_raw
        --- neural_textures_faces_c23
        --- neural_textures_faces_c40
        --- neural_textures_faces_raw
        --- original_faces_c23
        --- original_faces_c40
        --- original_faces_raw
    --- retina
        --- deepfakes_faces_c23
        --- deepfakes_faces_c40
        --- face2face_faces_c23
        ...
    c40_all.pkl
    data_all.pkl
    retina_all.pkl

Use script to get images' path pkl.

cd src
python get_train_val_data.py

Run train_resnext.py script.

cd src
python train_resnext.py