Prepare dataset

Training dataset mount to /code/train. Public test 2 dataset mount to /code/public_test_2. Public test 1 dataset mount to /code/public_test

Split videos to 2 folders fake and real python split_data.py

Extract all even frames from videos

python extract_frames.py --config extract_frames_train_config.json
python extract_frames.py --config extract_frames_pub1_config.json
python extract_frames.py --config extract_frames_pub2_config.json

Split k-fold

python split_kfold.py

Detect face

cd /code/yolov7-face-main
python detect.py

Training

CUDA_VISIBLE_DEVICES=0 nohup python code/train_teacher.py --exp exp_39 > nohup_exp_39.out &
CUDA_VISIBLE_DEVICES=0 nohup python code/train_teacher.py --exp exp_40 > nohup_exp_40.out &
CUDA_VISIBLE_DEVICES=0 nohup python code/train_teacher.py --exp exp_43 > nohup_exp_43.out &
CUDA_VISIBLE_DEVICES=0 nohup python code/train_teacher.py --exp exp_49 > nohup_exp_49.out &

Experiment result

Exp	Pub 2	Private
Ensemble exp_39, 40, 43, 49, B5, 3TTA, 11 frames	0.00855
Ensemble exp_39, 40, 43, 49, 2TTA, 8 frames	0.019
Ensemble exp_39, 40, 43, 49, 2TTA, 8 frames, fp16	0.02137	0.06373
Ensemble exp_40, 43, 49, B5, 2TTA, 11 frames, fp32	...	0.05882
Exp_39, 3TTA, 7 frames, best eer	0.02778
Exp_40, 3TTA, 11 frames, best acc	0.02991
Exp_40, 3TTA, 11 frames, best eer	0.02381
Exp_43, 3TTA, 11 frames, best acc	0.02381
Exp_49, 3TTA, 11 frames, best acc	0.02778
B5, 3TTA, 11 frames, best acc	0.02564

Submission

Gen submission csv file

python code/submission/submission.py

Ensemble

python code/submission/ensemble_multimodel.py

Visualize grad-cam

python code/visualize_grad_cam.py