Why is the generalization performance of the model I trained poor?

[Swilder7]

First of all, the results are as follows.

Celeb-DFv2:Current ACC, AUC, AP, AR, mF1 for ['Celeb-real', 'Celeb-synthesis', 'YouTube-real'] --- ['real', 'fake'] -- 51.68067226890757 -- 76.33165164484251 -- 84.69933774463259 -- 60.50110537951363 -- 70.58385564404473

c23:Current ACC, AUC, AP, AR, mF1 for ['Original', 'Deepfakes', 'Face2Face', 'FaceSwap', 'NeuralTextures'] --- ['real', 'fake'] -- 60.71428571428571 -- 91.63265306122449 -- 90.33479567064433 -- 60.71428571428572 -- 72.62027076235252

c0:Current ACC, AUC, AP, AR, mF1 for ['Original', 'Deepfakes', 'Face2Face', 'FaceSwap', 'NeuralTextures'] --- ['real', 'fake'] -- 97.14285714285714 -- 99.50510204081633 -- 99.4571011891637 -- 97.14285714285715 -- 98.28635829459391```

I'm not quite sure where the problem lies with the model I trained. I basically followed the steps in every aspect. My steps are as follows:

1. python /data3/yxy/LAA-Net-main/scripts/package_utils/images_crop.py -d Original -t train -n 128
2. python /data3/yxy/LAA-Net-main/scripts/package_utils/geo_landmarks_extraction.py --config /data3/yxy/LAA-Net-main/configs/data_preprocessing_c0.yaml --extract_landmark --save_aligned
3. python /data3/yxy/LAA-Net-main/scripts/package_utils/images_crop.py -d Original -c c0 -t val
4. python /data3/yxy/LAA-Net-main/scripts/package_utils/geo_landmarks_extraction.py --config /data3/yxy/LAA-Net-main/configs/data_preprocessing_c0.yaml --extract_landmark --save_aligned
5. CUDA_VISIBLE_DEVICES=0 python /data3/yxy/LAA-Net-main/scripts/train.py --cfg /data3/yxy/LAA-Net-main/configs/efn4_fpn_sbi_adv.yaml

This is my configuration file. efn4_fpn_sbi_adv.json

And there have been no changes made to the code in the file except for modifying the root.

I would be extremely grateful if you could spare some time to take a look at my situation and give me some insights or suggestions on what might have gone wrong. Your expertise and guidance would be of great help to me in solving this issue.

Thanks!

[Swilder7]

By the way, the model was trained from scratch instead of loading the provided pre-trained model during the training process.

[begin20231024]

By the way, the model was trained from scratch instead of loading the provided pre-trained model during the training process.

hi, can u share the rest test results to me ? such as DFD and DFDC? I have met the problem like u ! Thx !