arrowluo / grace Goto Github PK

您好，参照您的readme参数设置复现三种情况中的结果，但是发现加上VAT之后效果反而变差了，并且其余结果也均比论文中的效果低，请问是为什么呢？readme里面的参数有需要调整的吗？谢谢。

您好，我在您代码的基础上在bertSequenceLabeling的基础上init初始化的时候添加了层，发现即使forward中不使用效果也会变差，请问这个load原始模型的过程是哪里需要修改吗？添加新的层是会影响其它层load的参数吗？不是太理解，期待回复，谢谢

Please provide a license. Thanks!

Thanks for the pretrained_weight you provided.

Could you also provide code for gradient harmonized loss?

Currently, other repo's GHM-C codes can only be applied to Binary Class, causing trouble.

Hi

Thank you for good repo, I want train this model on completely new domain dataset, I have review comment, its aspects and polarity respectively. How do i generate the training and eval data set to use these models.

Token, POS, CHUNCK, Term Label, Sentiment Label, Collapsed Label

I could think of using spacy get token, pos and noun-chunks and sentiment label is what aspect sentiment, but what is term label, how can get this, my understanding is collapsed label - term label + sentiment label

Hi @ArrowLuo
Thanks for the implementation and I was able to run it successfully on the latest pytorch version (built from source), in order to use cuda 11.2.. and was able to replicate your paper results.

However I'm trying to create a script to use the trained model and make predictions only. Was wondering if you had some script or code that could help, as I was going through the code to patch together something..
Cheers.