littttttlebird

Results 5 issues of littttttlebird

![image](https://user-images.githubusercontent.com/20662579/76062085-ea266b00-5fbf-11ea-8c70-26571a740f6e.png) 在一般蒸馏过程中,会将soft_target以及hard_target都用上,构造两个交叉熵损失,为什么这里只用了soft_target来计算prediction layer的loss(或者说为什么hard_target cross entropy 权重为0)?

I have a text multi-label classification task,can i use supCon loss ? supCon loss is accumulated by every label view,for example: `batch data label = [[1, 0, 1], [0, 1,...

跑出来结果,解码结果总是空的

看咱们HMCN-F的实现代码,标签字典是通过数据集里面标签构造的,比如样本集如下: > case1 A case2 A--A1 case3 B--B1--B11 case4 B--B1 case5 B--B1--B12 那么将会得到标签字典如下: >A A--A1 B--B1 B--B1--B11 B--B1--B12 在训练的时候,一级类目标签有[A],二级类目标签有[A--A1, B--B1],三级标签有[B--B1--B11, B--B1--B12]。 以case3为例,真实标注标签为B--B1--B11 - 在分级标签预测上,一级标类目上标签是[0],二级类目上标签是[0, 0],三级类目上标签是[1, 0]; - 在全局标签预测上,真实类目标签是[0, 0, 0,...

For example, in a companion app, the user (alice) says to the assistant (bob):"Yesterday my good friend jack said that he didn't like playing football, but his father enrolled him...

question