PaperDiscovery
commented
1 year ago The train_first is our experiment version, so there are several other methods in that function that we don't use in final version, such as M times dropout mentioned in our section and you can see just as R-drop, we use KL divergence.
我的理解是调用train_simple_model函数完成后直接对contrastive_network进行训练，这个train_first函数的作用是什么？并且这个函数求kl散度，论文也没有提及。