The AUC of CelebDF
JieLiu95 opened this issue · comments
I run the inference on Celeb-DF-v2 with your pre trained model, but the AUC is 0.9235, which is lower than yours. And the versions of torch and torchvision are the same with you. Do you know the reason? Can you provide the version of opencv?
The version of opencv (opencv-python) is 4.5.4.60.
It is strongly recommended to use the provided docker file to reproduce the experimental results.
Thanks a lot. I have solved it when I change the version of efficientnet-pytorch.
@JieLiu95 Can you provide the version of efficientnet-pytorch!?
0.7.1 @YU-SHAO-XU
@JieLiu95 and your SBI (train) is 1421 !? i am sure that landmark augmentation is used because show 'exist library' . I train from scratch and test on CDF AUC only 90.24, so i wanna check if development environment or datst have problem thank you!
@JieLiu95 did u train from scratch ? or only use weight that author provide?!
@JieLiu95 and your SBI (train) is 1421 !? i am sure that landmark augmentation is used because show 'exist library' . I train from scratch and test on CDF AUC only 90.24, so i wanna check if development environment or datst have problem thank you!
I train it from scratch. Did you utilize the mask of X-ray as illustrated in Step 4.
mkdir src/utils/library
git clone https://github.com/AlgoHunt/Face-Xray.git src/utils/library
@YU-SHAO-XU
@JieLiu95
is that bi_online_generation.py file ? if use that file AUC will 77.xx increase to 90.xx, or you mean DeepFakeMask.py !? and how to use DeepFakeMask !?
If you use bi_online_generation.py, I guess you have used DeepFakeMask.py by the function random_get_hull.
@JieLiu95 so i think it seems to dataset or development environment ~ thank you !
@JieLiu95
did u hvae error like :
- (-215:Assertion failed) !ssize.empty() in function 'resize'
- [W pthreadpool-cpp.cc:90] Warning: Leaking Caffe2 thread-pool after fork. (function pthreadpool)
is it normal !?