search

1024er / cbert_aug

68 stars 19 forks source link

关于论文实验的做法的问题 #2

Open GW-S opened 4 years ago

GW-S commented 4 years ago

你好，我想咨询一下，在代码中，每一个epoch的训练，都会产生一个新的增强数据集，假设在第三个epoch中，增强的数据使得RNN+SST5的效果变好了，那在第四个epoch中会不会导致RNN+SST5的训练效果变差。您有做过实验证明，3epoch的效果好，那么4，5，6，epoch的效果一定好的实验吗？

mmmjung commented 4 years ago

Hi, It would be more helpful if you could describe the issue in English. Your issue might also be useful for others.

GW-S commented 4 years ago

Hi, It would be more helpful if you could describe the issue in English. Your issue might also be useful for others.

OK,I will do this. hello,author,I want consult you a question,in your paper and your code, we can see each epoch,you will augmentation one fine_tuning model , and you can use this fine-tuning model augmentate a new dataset, but we can see ,in epoch 0, the augmentation data will be useful, but in epoch 1,the augmentation is not useful ,is that a normal thing,or just a random augmentation? I think you should talking about that thing.

1024er commented 4 years ago

Sorry for not being able to respond in time due to busy work.

I think it may happens, but we suggest to evaluate the model on validation data at the end of each epoch, then taking the model which has the best validation performance and evaluating it on test data.

There was recently a post which compared CBERT augmentation with other methods, even in extremely low resources situations. I think it may helps you to some extent.

https://mp.weixin.qq.com/s/ZmnP321nq_7wX4_HHskv_w

GW-S commented 4 years ago

Thank you very much.  The link you give me is helpful. 

------------------ 原始邮件 ------------------ 发件人: Xing Wu <notifications@github.com> 发送时间: 2020年3月23日 09:39 收件人: 1024er/cbert_aug <cbert_aug@noreply.github.com> 抄送: 盛国威 <945786604@qq.com>, Author <author@noreply.github.com> 主题: 回复：[1024er/cbert_aug] 关于论文实验的做法的问题 (#2)

Sorry for not being able to respond in time due to busy work.

I think it may happens, but we suggest to evaluate the model on validation data at the end of each epoch, then taking the model which has the best validation performance and evaluating it on test data.

There was recently a post which compared CBERT augmentation with other methods, even in extremely low resources situations. I think it may helps you to some extent.

https://mp.weixin.qq.com/s/ZmnP321nq_7wX4_HHskv_w

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.