FrankWork / fudan_mtl_reviews

TensorFlow implementation of the paper `Adversarial Multi-task Learning for Text Classification`
178 stars 40 forks source link

domain_loss的一些问题 #16

Open bobobe opened 5 years ago

bobobe commented 5 years ago

您好,在我们的代码中,利用gradient reverse得到对抗损失,但是这个损失一直不变稳定在0.7左右,显然是对抗不起作用,而作者您的实验结果中domain_loss是从0.6下降到0.2,我想问下这是什么问题,对抗训练有什么技巧吗?谢谢