tsujuifu / pytorch_graph-rel

A PyTorch implementation of GraphRel

Tsu-Jui (Ray) Fu

I am a research scientist at Apple AI/ML. My research lies in vision+language and text-guided visual editing. I obtained my Ph.D. in computer science from UC Santa Barbara, advised by William Yang Wang. I am also interested in video summarization and information extraction. My goal is to bridge the gap between multiple modalities via the AI system.

pytorch_graph-rel's Issues

about some details

Hi,
How can I express it by ans_rel / wgt_rel for the case of entity pairs overlapping?

code

可以提供这篇论文的源代码吗，多谢了

Dataset format

Hello, I want to reproduce the code but I encounter a problem. What's the meaning of the parameter idx, inp, pos, etc. in the following code

for idx, inp, pos, dep_fw, dep_bw, ans_ne, wgt_ne, ans_rel, wgt_rel in ld_ts:
print(idx.shape)
print(inp.shape, pos.shape, dep_fw.shape, dep_bw.shape)
print(ans_ne.shape, wgt_ne.shape)
print(ans_rel.shape, wgt_rel.shape)

about weight_loss

Hi, I 'd like to know why you set weight _ loss to 2?

i reproduced the data preprocessing

I reproduced the data preprocessing, but the loss cannot be reduced, and the model could only predict "O" about entities and "No relation" about relations. The preprocessing and framework of the model have been written, and the forward propagation of the model used the source code in ipynb given by Fu. Welcome to my repository to discuss with me.
https://github.com/ShayneCheng/GraphRel

Why don't release the preprocessed data?

Without your preprocessed data and the format of your data, what's the difference between releasing codes and not releasing?

Could you tell me what does -1 means in wgt_ne and wgt_rel?

数据集下载

你好呀，我是关系抽取领域的新人，请问数据集NYT和WebNLG有相应的下载链接咩~谢谢

I have reproduced a dataset in JSON format with the following code, which is not guaranteed to be the same as the author's implementation.

pkl dataset

Can you provide the dataset from json format to pkl？ I couldn't find the relevant codes.
Thanks a lot!

pretrain process

Hi, guys. Can anyone process the dataset from json format to pkl？ I couldn't find the relevant codes.

data processing

Thank you very much for publishing the paper code. Can you update the code of the data processing module?

How do you predict new data and interpret the model outputs?

Sorry I'm the new in this field and I checked the code, but I want to know that how do you predict new data and interpret the model outputs? I'm kind of confused.

Where can I get this pre-processed file abount pre_tr.pkl,Please.

Where can I get this pre-processed file abount pre_tr.pkl,please.

I reproduced the dataset preprocessing script. And this model cannot be reproduced.

Here is my script: script
I rewrite the whole model and find that it always predict 'O' for entity and 'no-relation' for relationship. The gold label for entity and relation is too sparse and the model is very difficult to train.

anybody knows the differences between tr.pkl vl.pkl ts.pkl and pre_tr.pkl pre_vl.pkl pre_ts.pkl

When I tried to reproduce this paper, I found that six files were read in the source code, namely tr.pkl vl.pkl ts.pkl and pre_tr.pkl pre_vl.pkl pre_ts.pkl.
I know the format of files beginning with the name pre; but what about the first three files? Does anyone know how it is generated?

Preprocess Dataset Code

Hi
Can you share your preprocess data code ? Thank you !

Original Dataset

Hello! Can you share nyt and webnlg original dataset containing train, dev, test ? Thanks a lot !

The resources are dismissed now

All the resources are dismissed now, where and how can I find them?

about dataset

hello!
I'm reproducing your code but I cannot get the loss convergence. I'm wondering if the dataset was preprocess in a wrong way. So I want to know the format of "pre_train.pkl".

有人实现中文数据集的处理吗？

比如阿里天池的CMeIE，中文用spcay 做分词逻辑和英文的不一样呢

[2021/12/20 Update] Reimplementation by Author

[2021/12/20 Update]
Thanks to everyone's interest in this project and sorry for missing the original preprocessed data.
It got lost in my previous lab, and I finally had time to reimplement it 😂.
I also want to appreciate @LuoXukun for his nice reply about reproducing.

NYT	Precision	Recall	F1
GraphRel1p (Paper)	62.9	57.3	60.0
GraphRel1p (Reimplementation)	60.9	59.2	60.1
GraphRel2p (Paper)	63.9	60.0	61.9
GraphRel2p (Reimplementation)	63.1	60.2	61.6

Hello, I would like to know where the three learnable relationship matrices are represented in the code

Hello, I would like to know where the three learnable relationship matrices are represented in the code

Has anyone implemented and processed WebNLG datasets ？

The author only provides the preprocessing of NYT data set. I would like to ask if anyone has implemented the preprocessing of WebNLG dataset and could you share it

There is a small BUG in the construction of the dep_bw(backward dependency parse tree)

pytorch_graph-rel/dataset.py

Line 42 in 9f0697d

dep_fw[i], dep_bw[i] = dep_fw[i]/sum(dep_fw[i]), dep_bw[i]/sum(dep_bw[i])

The above regularization method does not regularize dep_bw on a per line basis, I think it would make more sense to change it to the following:

        for i, w in enumerate(res):
            inp_sent[i], inp_pos[i] = w.vector, self.POS[w.tag_]
            
            dep_fw[i][i], dep_bw[i][i] = 1, 1
            for c in res[i].children:
                for j, t in enumerate(res):
                    if c==t:
                        dep_fw[i][j], dep_bw[j][i] = 1, 1
            # dep_fw[i], dep_bw[i] = dep_fw[i]/sum(dep_fw[i]), dep_bw[i]/sum(dep_bw[i])
        dep_fw[:i+1] = dep_fw[:i+1] / dep_fw[:i+1].sum(axis=-1, keepdims=True)
        dep_bw[:i+1] = dep_bw[:i+1] / dep_bw[:i+1].sum(axis=-1, keepdims=True)

entity pair overlapping

How do you deal with the entity pair overlapping cases when using 'softmax' in your model ?

Please provide dataset preprocessing details or the structure of pkl file for training?

Equations and Code Matching Issue

Hi,

First, i cannot figure out what is the real meaning of h_w1 and h_w2 in the paper. author never gives a description of them based on the below equation.

then, I start to search them in code, and I find they are totally same. please correct me if I'm wrong. the calculation of S(w1, r, w2) uses the below code.

however, both trs0_rel and trs1_rel are linear functions with same dimension. same input are passed to them and i think the output should be also same.

self.trs0_rel = nn.Linear(self.hid_size2, self.hid_size)
self.trs1_rel = nn.Linear(self.hid_size2, self.hid_size)

so my question is, if one linear function is enough, why two are listed with different notations?

any explanations are greatly appreciated. thank you

Anyone reproduce this paper?

Could you provide me data? I try to reproduce it but fail.

code details

I didn't get the data in pkl format, so I want to ask some details in the code:
1.dataloader:
for idx, inp, pos, dep_fw, dep_bw, ans_ne, wgt_ne, ans_rel, wgt_rel in ld_ts
what are the meanings of these variables?
2. pos is or not pre-characterized to the sentence.Is this part of speech tagging?
3. The adjacency matrix in the GCN should obtain the dependency of the related words. Is this the syntax dependency obtained by using the spacy?

code

可以提供这篇论文的源代码吗，多谢了