search

DariusAf / MesoNet

"MesoNet: a Compact Facial Video Forgery Detection Network" (D. Afchar, V. Nozick) - IEEE WIFS 2018
Apache License 2.0
252 stars 114 forks source link

Method to get weight #31

Closed Uhouhzam closed 2 years ago

Uhouhzam commented 2 years ago

Hello, I am a student of master. I work for my research project deepfake detection. I wanted to test your code with other deepfake dataset. I used your code Meso4 and MesoInception to test other deepfake dataset. It doesn't work as better as writed in the papier. I think it's problem of weight. Can I know what kind of method that you used to get your weight, please. Did you use backpropagation to get your weight and save them as file ?

DariusAf commented 2 years ago

Backpropagation would give you the gradient of the weight, not their value. It was trained in a very standard way that is described in the paper. And saved using tf.keras.Model.save_weights() to the H5 format, available on this repo.

And yes, the method was one of the first paper on deepfake detection, the scope was simply to train on what we collected as also one of the first deepfake dataset. It was never claimed that it works out-of-distribution or out-of-task, though this is an important aspect that has been studied by subsequent papers.

Uhouhzam commented 2 years ago

Thank you for your answer . It's very helpful for me.