Binary FFPP structure dataset folder

[itlvd]

Hi,

Can you tell me about the structure of binary classification in the FFPP dataset?

My structure has two folders: Real (origin) and Fake (4 attack methods).

And what happens when I use two folders for two classes for training with train_binary_ffpp.py and testing with test_binary_ffpp.py

Thank you.

[honghuy127]

Hi,

The code uses the standard torchvision.datasets.ImageFolder class. As its mechanism, you can add the prefixes 0_ and 1_ to force it to map these folders to [0, 1] labels in order.

You could organize the dataset as follow:

+ FFPP (--dataset)
   + Train (--train_set)
       - 0_Real
       - 1_Fake
   + Val (--val_set)
       - 0_Real
       - 1_Fake
    + Test (--test_set)
       - 0_Real
       - 1_Fake

Since the code can convert any label > 1 to 1, you can go further like:

- 0_Real
- 1_DF
- 2_F2F
- 3_FS
- 4_NT