This project fork from https://github.com/ondyari/FaceForensics.

It is convenient to implement their trained model by this original project.

This forked DIY version provides TOOLS for:

In directory [/FaceForensics/DataPreprocess]:

• [ExtractVideoFrame.py] : Extract frame from provided video dataset.

• [ExtractFaceFromFrame.py] : Extract face from every frame. Because the paper introduce a face detection algorithm to enhance the performance of fake detection.

• [SplitDataset.py] : Split dataset into training, validation and test set. The ratio mentioned in their paper is 0.72. 0.14, 0.14.

• [MergeDataset.py] : Merge dataset of manipulated images(Deepfake, Face2face, Faceswap and NeuralTextures) into type 1 and original images into type 0.

In directory [/FaceForensics/TrainModel]:

• [TrainModel_2.py] : Use a rransfer learning template in pytorch.
  • Choose your merged dataset path in Line 25
  • Choose batch size in line 37
  • Pretrain the last linear layer with 3 epochs by adding parameter '----pretrain3epochs' like TrainModel_2.py --pretrain3epochs
  • Train whole model with 15epochs directly TrainModel_2.py
  • Save the trained model in Line 199(specify your output path and filename here).

You can see the testing and validation accuracy in the console during training time. If you want to see the test set counterpart, use code in directory [/FaceForensics/TestModel]:

• [test_model.py] : Detect all images in a directory and write all prediction output into a .txt file.
• [test_model_GUI.py] : Detect single video and output a series of frames with face box and prediction tag.
• [test_our_model.py] : Implement our trained model of [TrainModel_2.py] and output prediction tag into .txt file. Implement format is : test_our_model.py -i /test_set_images_path -m /model_path -o /output_txt_file_path

After get prediction output by [test_our_model.py], use code in directory [/FaceForensics/StatResult] to stat accuracy:

• [cal_accuracy.py] : stat whole test set accuracy.
• [cal_accuracy_single.py] : stat single test set accuracy, eg. Deepfakes.

If you have any questions, contact me via [email protected].