thuml / debiased-self-training Goto Github PK

我看代码仓里给的那个好像是67.8.但是训练的参数好像和论文里面不太一致。70.4的这个参数是如何设置的呢老哥。我看其他论文好像有些是训练2^20轮，咱们这个也是需要这么设置嘛。

Hi thank you for the great work! I am now trying to use DST on my own dataset. When training the model, the cls_loss and worst_case_loss decrease normally, but the self training loss always increase, the accuracy seems ok but I don't understand why self training loss increase. Is this behavior expected?

大佬您好！非常感谢您工作对于我的启发，关于梯度反转中max_iter的大小对于训练结果的影响，还请您指教一下；另外我注意到您之前的回答提到对抗中两个player的balance很重要，请问衡量这个balance有什么标准吗？
感谢！