thecharm / mega Goto Github PK

咨询一下，这个模型训练使用的显卡设备是什么型号，想要做多模态关系抽取，至少需要什么型号的显卡

您好，我下载了数据集，但是图片无法查看，请问需要用什么软件打开图片。

请问一下，数据标注的时候是对一个句子中的所有实体对都会进行标注关系吗，因为观察到数据集中有大量的关系为None。

请问论文里代码结果随机种子的设置是多少呀？目前无法复现出论文的结果？其他设置与github描述设置的一致，但是设置不同的随机种子结果差异特别大，请问论文中的结果设置的随机种子是多少呀？

Dear developers hello ~
I need help！！
What is the meaning of the 'RT' field in the TXT file in the dataset？？
I can't understand this, but it's in every line55555

您好，我是Mega这篇论文的忠实的读者，我想问一下这个alignments.py里面 get_estimated_alignment_matrix函数为啥没有代码实现？
Hello,I am a reader of this paper.I want know that why this function called "get_estimated_alignment_matrix" in the file of "alignments.py" have no code implementation.

Hi, nice work and nice repo, but it seems that this repo does not contain the graph alignment part?

你好，我下载的ckpt里面是MEGA_ckpt.tar压缩包，并不是ckpt，而且训练的话，效果好像达不到论文里面的准确率

Hi, here are some questions puzzling me:

1. There is one file named "rel_1" in the dataset.zip. What information does it contain? Does it represent the aligned maps between the text-graph and visual-graph? Btw, I noticed that the documents in it have different lengths.
2. There are items named "selected objects" / "objects relations" in the case study table. Do these items represent the most aligned visual objests or the relations between these aligned visual onjects?

您好，加载您提供的ckpt模型时，出现报错

RuntimeError: Error(s) in loading state_dict for SoftmaxNN:
Missing key(s) in state_dict: "sentence_encoder.bert.embeddings.position_ids".

Hello, could you please provide the following pdf version of this paper? I want to read this paper. My email is [email protected]. Thx.

我已经依据脚本下载了数据集，分别是fewrel、nyt10、semeval和wiki80。这几个数据集没有图像数据如何利用这些数据去运行程序？还需要预处理这些数据？

请问有人遇到这个问题吗？最后一个batch的pic size是 torch.Size([58, 4096])

Dear developers, I have encountered some more difficult problems 5555~
The project always shows that the path does not get the expected value, I do not understand how to solve this, very much looking forward to your reply! Thank you very much!

The error code is as follows：

self.pic_file_list = os.listdir(self.pic_path)
for pic_file in self.pic_file_list:
    with open(os.path.join(self.pic_path, pic_file), 'r', encoding='utf-8') as f:
        line_list = f.readlines()
        feature_list = line_list[2].strip().split('\t')
        class_list = line_list[1].strip().split('\t')

这里的pic_path，pic_file分别是imgSG/train， twitter_19_31_0_0.jpg时 出现'utf-8' codec can't decode byte 0xff in position 0: invalid start byte错误
想请教下读取图片特征时应该如何修改以及 class_list 和 feature_list分别是什么？

您好，祝贺您中了MM，我想看一下这个论文的内容，不知道您是否可以给我发一份pdf呢？我的邮箱是 [email protected]，非常感谢！

作者你好，很高兴能够看到你的开源模型，在学习你的模型中关于图对齐部分，在StructuralAlign.py中的path_load应该写什么呢，我用的是data/combined_graph_edges_10/train，以及在定义属性attributes时，涉及到存节点属性的numpy矩阵文件路径，这个路径是什么？我没有找到，期待你的答复！