hanjiale / hcrp Goto Github PK

Hello, thanks for your interesting work. However, I encountered some problems with reproducing results on the validation set. I did not modify the parameters in the code and just run 'python train.py'. The results I obtained on the validation set 'val_wiki.json' and the 5-way 1-shot setting, which should be the FewRel 1.0, were just about 84%. There is a great gap between my reproducing results and your provided results "90.9%". Thus, I want to verify

1. whether the parameters you used for training are the same as the ones in this code;
2. did you just use the original BERT for training without external pertaining.

Looking forward to your answer, thanks!

作者您好！ 感谢您出色的实验，这里有一个有关模型推理的问题：

您的模型实现中，模型推理时需要传入rel_text，即关系类型（及其描述），fewRel的测试集中，support set并没有给出每个support instance的关系类型，想请问一下，您在测试集上做推理时，是如何处理support set对应的关系类型的呢？

# in model.py

def forward(self, support, query, rel_text, N, K, total_Q, is_eval=False):
        """
        :param support: Inputs of the support set. (B*N*K)
        :param query: Inputs of the query set. (B*total_Q)
        :param rel_text: Inputs of the relation description.  (B*N)
        :param N: Num of classes
        :param K: Num of instances for each class in the support set
        :param total_Q: Num of instances in the query set
        :param is_eval:
        :return: logits, pred, logits_proto, labels_proto, sim_scalar
        """

Hi @hanjiale,

Thank you very much for releasing the source code of your great work.
I see your source code is very clear and easy to follow. I only wonder about calculating the CE loss at this line.
If possible, Could you please explain why we need to get the min value of logits and then get the max value on N+1 values? As I think we can only simply get the max value on N values...
I am looking forward to hearing from you. Many thanks!

请问对比RPCL的可视化是怎么实现的，需要保存什么样的数据？