yingrliu / taan-mtl Goto Github PK

Dear author：
dataTool.MAIN_DIR目录下的文件结构是什么样的呢？

It seems that there is something wrong about your function "LossCosine", located at /Layers/TAL_pytorch.py. When I train with cosine regularization, the loss is less than zero, which is totally impossible. And I notice that in function "LossCosine", before returning the mean value of loss, you multiply it with minus one, which is different from that in function "LossDis" function.
And in your paper, the math formulation of your Cosine regularization is also multiplied by minus one, which makes the loss value is less than zero. Is there something wrong with me?

首先非常感谢你的非常优秀的工作。

其次在拜读了你的文章和复线你的代码之后，我有以下几个疑问，希望能得到你的答复。

1. 为什么文章中对比方法中没有mmoe模型呢？我看你的代码中已经实现了mmoe方法（TAAN 结果要明显优于mmoe文章汇报的结果）

2. 我在自己的业务场景下跑了TAAN的网络，在我的数据集场景下，结果震荡比较严重。请问这是正常的么？你之前有遇到过么？对于结果震荡你有什么调参的建议么？