vincent-zhq / mrdf Goto Github PK

conda create -n df116 python=3.8
pip install -r requirements.txt

-- dataset (FakeAVCeleb)
-- model: 
    avhubert (download from https://github.com/facebookresearch/av_hubert/tree/main/avhubert) with some small changes, ImageEncoder.py, model.py
    model realted files (__init__, avdf, avdf_ensemble, avdf_multiclass, avdf_multilabel, mrdf_margin, mrdf_ce)
-- fairseq: (download from https://github.com/facebookresearch/fairseq with some small changes)
-- outputs: log, results, ckpts
-- data/FakeAVCeleb_v1.2: path to the dataset (FakeAVCeleb) and splits (5 folds)
-- utils: loss, figure, utils
-- main.py
-- train.py
-- test.py
-- requirements.txt

## train example
CUDA_VISIBLE_DEVICES=0 python train.py \
  --model_type MRDF_CE --save_name MRDF_CE \
  --data_root ./data/FakeAVCeleb_v1.2/ \
  --dataset fakeavceleb

CUDA_VISIBLE_DEVICES=0 python train.py \
  --model_type MRDF_Margin --save_name MRDF_Margin \
  --data_root ./data/FakeAVCeleb_v1.2/ \
  --dataset fakeavceleb

## test example
CUDA_VISIBLE_DEVICES=0 python test.py \
  --data_root ./data/FakeAVCeleb_v1.2/ \
  --checkpoint /Path/To/MRDF_Margin_train_2.ckpt \
  --train_fold train_2.txt --model_type MRDF_Margin

CUDA_VISIBLE_DEVICES=0 python test.py \
  --data_root ./data/FakeAVCeleb_v1.2/ \
  --checkpoint /Path/To/MRDF_CE_train_2.ckpt \
  --train_fold train_2.txt --model_type MRDF_CE

@article{zou2024cross,
  title={Cross-Modality and Within-Modality Regularization for Audio-Visual DeepFake Detection},
  author={Zou, Heqing and Shen, Meng and Hu, Yuchen and Chen, Chen and Chng, Eng Siong and Rajan, Deepu},
  journal={arXiv preprint arXiv:2401.05746},
  year={2024}
}