nju-websoft / adalogn Goto Github PK

No such file or directory: 'ReclorDataset/ReclorDataset_EDUs.json'

hello, very valuable work, I train from sctrech, but I find it is very time costing.
batch size is 8, but one step need 13 min!

0%| | 1/2315 [13:20<514:50:48, 800.97s/it]

bert and dgl all in the gpu, 3080ti 12GB. Is it normal for this model?

Hi authors,

Could you please tell me which mode that you used (server/cli) and how can i create FLAT files like you.

Thanks for you support.

作者您好，复现该项目代码之后，有两个疑问。
第一，我使用了您在谷歌云盘上传的checkpoint对两个数据集进行了验证。其中，LogiQA的验证集和测试集、Reclor的验证集结果基本符合您的论文数据。对Reclor的测试集，我给test.json补充了标签，得到的测试结果是74.4%，远高于论文数据。
第二，我尝试在3块RTX8000上从头开始训练模型，参数和原代码中的一致。但是训练后模型在验证集与测试集上的正确率远低于预期，似乎loss也没有收敛。请问参数需要更改吗？
祝您新年快乐，工作顺利！

reclor数据集没有测试集的标签，请问如何论文中reclor的test-acc如何计算得出

你好～
我在尝试运行eva的时候，出现以下错误：
RuntimeError: Error(s) in loading state_dict for Tagger:
size mismatch for drop_replacement: copying a param with shape torch.Size([2248]) from checkpoint, the shape in current model is torch.Size([325]).
希望能得到解答，谢谢～

Hello, thanks for your work !
Is there something wrong in class RGAT's forward function of GNNs.py ?

In Section2.3.2 of the paper, (l+1)-th iteration is relative directly with l-th iteration. But in the code, I do not see any connections that output in each iteration is individual. Specifically, there are some local variables outside in the code. Should they be put inside the for ?

您好，可以给出您使用Graphene提取工具的graphene-core / graphene-server / graphene-cli中哪个模式和使用的命令吗？

transformers==4.5.0
module transformers has no attribute TFRobertaAdaLoGN
which version?

我不是很懂，因为课程原因想运行一下您的代码
在服务器上运行时好像单个显卡运行时显存不够（11G显存），想修改batch size，在shell里面将两个batch size都修改为了1还是跑不动。
用两块以上的显卡一起跑的时候会有RuntimeError: NCCL Error 2: unhandled system error报错，不知怎么解决，万分感谢！！