zhy

> Same here guys. Gets completely stuck. I tried dowgrading to 1.9.2 as suggested and it did not work. hello how are you deal with it? please help me

> After debugging, I found that when this error occurs, the response format returned by the server is different from the normal response format hello, how to solve??

已经解决了，在调用训练时候，训练集和验证集是不同的线程。所以你需要开启两个线程，-num_worker 2；然后在代码中写实例化两个对象bc1 = BertClient()，bc2 = BertClient()；线程很重要

好像明白了，最后的输出层的attention，没有再去和V相乘，对吗？

谢谢你，

> I meet same problem when use keras_contrib 2.0.8, tf 1.12, keras 2.2.5. After install keras 2.2.4, this error disappear. hahahah,,, before have a keras bug when keras 2.2.4,so i...

我已经进行了finetun，但是评估的时候现在seqeval.metrics和sklearn.metrics两个库的f1值不一样，sklearn的f1值高，可以达到96多且intent达到97%，但是seqeval的f1值只有91，我认为seqeval的对的，因为还有个指标sent一直都是81，所以应该是哪里不对。但是现在我已经死调烂调都是这个。

https://github.com/monologg/JointBERT/blob/7497631c20/data/vocab_process.py，这是论文对应代码，就是已经微调，结构bert+crf，但是怎么调都不能到达人家的精度