YOLO

> rainer.py第395行和第461行位置加上`prev_hidden_states = hidden_states` 好的，谢谢！ 请问这个是因为这个项目的作者写错了吗？导致一直生成重复的句子？加上这两行之后，才是正确的代码吗？ 还是这两行是对原模型的改进呀？

> I mean you can send e-mail. @cainiaoshiwo hello，the link is broken，could i get a new one？thank！

> > > rainer.py第395行和第461行位置加上`prev_hidden_states = hidden_states` > > > > > > 好的，谢谢！ 请问这个是因为这个项目的作者写错了吗？导致一直生成重复的句子？加上这两行之后，才是正确的代码吗？ 还是这两行是对原模型的改进呀？ > > 加上才是LSTM自回归解码的标准过程吧，不然解码器每次循环输入的隐状态变量都是一样的，会造成输出也是一样的 你好，抱歉再次打扰，还有个问题想请教一下。代码中不管是trainer中还是tester中，在生成句子和单词的时候都使用了 `for sentence_index in range(captions.shape[1])`和`for word_index in range(1, captions.shape[2])` 其中，captions是报告的groundtruth，而在test的时候我们是没有groundtruth的，这边应该怎么理解呀？

> Hi Liqq1, I also have the same question about this, do you have a explanation or suspect for this? Thansk a lot. Sorry, I haven't caught on yet. Are...

> Hi, thanks for your interest. As for the data split, we follow the same data split as [R2Gen](https://github.com/cuhksz-nlp/R2Gen) and [R2GenCMN](https://github.com/cuhksz-nlp/R2GenCMN) to randomly split the data (70-10-20). I have updated...

> > > Hi, thanks for your interest. As for the data split, we follow the same data split as [R2Gen](https://github.com/cuhksz-nlp/R2Gen) and [R2GenCMN](https://github.com/cuhksz-nlp/R2GenCMN) to randomly split the data (70-10-20). I...

> > > Hi, thanks for your interest. As for the data split, we follow the same data split as [R2Gen](https://github.com/cuhksz-nlp/R2Gen) and [R2GenCMN](https://github.com/cuhksz-nlp/R2GenCMN) to randomly split the data (70-10-20). I...