Mryangkaitong

这里都是笔者之前的一些demo用以入门，目前主要做深度学习方面的项目，这里就没有上更多有深度的项目了，后续有机会会更新。

> > > > 请问在ChineseNRE/data/people-relation/data_util.py的54行 > > set_ids = range(1, len(set_words)+1) > > 那么“所有单词”列表的id索引是从1开始的，而在使用nn.Embedding时，取某个单词的vec索引是从索引0开始的，不就混乱了吗？ > > 为什么不直接这样set_ids = range(len(set_words))写呢？ > > 谢谢 > > 0应该是用来给unknown word了，在后面找一下 第61行程序 id2word[len(id2word)+1]="UNKNOW" 不是将最后一个id号设为unknown word了吗？

> > > > 请问在ChineseNRE/data/people-relation/data_util.py的54行 > > > > set_ids = range(1, len(set_words)+1) > > > > 那么“所有单词”列表的id索引是从1开始的，而在使用nn.Embedding时，取某个单词的vec索引是从索引0开始的，不就混乱了吗？ > > > > 为什么不直接这样set_ids = range(len(set_words))写呢？ > > > > 谢谢...

> 非word应该是为了凑batch在后面添加的padding吧。 > > 代码你可以改一改试一下效果，我实在不记得代码细节了。 @Mryangkaitong 好的，谢谢

> > > sts-b任务的数据是有label的，为什么是无监督训练呢？不解 https://blog.csdn.net/weixin_42001089/article/details/109499760

> I have a similar question that why there is a peak in "loss/total" and "loss/policy"？ Is your kl becoming more and more negative? According to my understanding, the method...

Thanks a lot, it worked

> Note that the KL-divergence (`objective-kl`) should never be negative. This can happen if you generate text with settings that are not pure sampling (e.g. early stopping strategies, minimum generation...

> Note that the KL-divergence (`objective-kl`) should never be negative. This can happen if you generate text with settings that are not pure sampling (e.g. early stopping strategies, minimum generation...

> > > Note that the KL-divergence (`objective-kl`) should never be negative. This can happen if you generate text with settings that are not pure sampling (e.g. early stopping strategies,...