MINTIME-Multi-Identity-size-iNvariant-TIMEsformer-for-Video-Deepfake-Detection
MINTIME-Multi-Identity-size-iNvariant-TIMEsformer-for-Video-Deepfake-Detection copied to clipboard
davide-coccomini
About the question for poor performance
Hello, you guys did a very good deepfake video detection. But if you encounter some questions, I would like to ask you to help answer them: I successfully ran through, but at the cost of modifying several parts in it
(1. I added torch.backends.cudnn.enabled = false in front of the file. 2, I removed the state_dict in the state_dict = torch.load(pretrain_path),
and the code I run is "python predict.py --video_path examples/fake_1_face_0.mp4 --model_weights models/MINTIME_XC_Model_ checkpoint30 --extractor_weights models/MINTIME_XC_Extractor_checkpoint30 --config config/size_invariant_timesformer.yaml"。 When I use your examples' video test, the results are all tested really, please ask me what to do, what caused it, thank you for your patience
Hi, are you sure that the weights are loading correctly? Also, I understand why you removed "state_dict" but why did you disabled cudnn?
