Open Glisten-5481 opened 9 months ago
Thanks a lot!
I'm sorry to bother you again, but I wanted to confirm whether the file used for training on new tasks is TaskN_KD.py? It seems that the file train_CNN_SupCon_and_CE.py only includes the two loss functions mentioned in the article
Additionally, I would like to know if testing on the new datasets is conducted separately? For example, testing the metrics for DFDC-P and DFD independently. Thank you once again for your assistance!
Sorry, this "train new model by file 'train_CNN_SupCon_and_CE.py'" is my typo. You should train the new task model by using file 'TaskN_KD.py'.
Testing the metrics should to be independent for DFDC-P and DFD. For example, if you use 'FF++' as your task1 dataset, you will get a 'model1'. Then you use DFDC-P as task2 dataset to train model1, you will get model2. You should use test model2 on 'FF++' testset and get acc and auc. Then test model2 on DFDC-P testset get another acc and auc.
Thank you very much for following our work.
Firstly, you should download all datasets including(FF++,DFDC-p,CDF,DFD).
Secondly, you can use file 'train_CNN_SupCon_and_CE.py' to train your first detection model with FF++ dataset.
Then, you can use 'get_feature.py' , 'get_image_info.py' and 'create_memory.py' to construct your memory set.
Finaly, you could randomly pick up 25 train video in your new dataset and add them into your memery set to train new model by file 'train_CNN_SupCon_and_CE.py'.