lee-zix / re-gcn Goto Github PK

This is the official code release of the following paper: Zixuan Li, Xiaolong Jin, Wei Li, Saiping Guan, Jiafeng Guo, Huawei Shen, Yuanzhuo Wang and Xueqi Cheng. Temporal Knowledge Graph Reasoning Based on Evolutional Representation Learning

Python 100.00%

re-gcn's Introduction

Temporal Knowledge Graph Reasoning Based on Evolutional Representation Learning

This is the official code release of the following paper:

Zixuan Li, Xiaolong Jin, Wei Li, Saiping Guan, Jiafeng Guo, Huawei Shen, Yuanzhuo Wang and Xueqi Cheng. Temporal Knowledge Graph Reasoning Based on Evolutional Representation Learning. SIGIR 2021.

Quick Start

Environment variables & dependencies

conda create -n regcn python=3.7

conda activate regcn

pip install -r requirement.txt

Process data

First, unzip and unpack the data files

tar -zxvf data-release.tar.gz

For the three ICEWS datasets ICEWS18, ICEWS14, ICEWS05-15, go into the dataset folder in the ./data directory and run the following command to construct the static graph.

cd ./data/<dataset>
python ent2word.py

Train models

Then the following commands can be used to train the proposed models. By default, dev set evaluation results will be printed when training terminates.

Make dictionary to save models

mkdir models

Train models

cd src
python main.py -d ICEWS14s --train-history-len 3 --test-history-len 3 --dilate-len 1 --lr 0.001 --n-layers 2 --evaluate-every 1 --gpu=0 --n-hidden 200 --self-loop --decoder convtranse --encoder uvrgcn --layer-norm --weight 0.5  --entity-prediction --relation-prediction --add-static-graph --angle 10 --discount 1 --task-weight 0.7 --gpu 0

Evaluate models

To generate the evaluation results of a pre-trained model, simply add the --test flag in the commands above.

For example, the following command performs single-step inference and prints the evaluation results (with ground truth history).

python main.py -d ICEWS14s --train-history-len 3 --test-history-len 3 --dilate-len 1 --lr 0.001 --n-layers 2 --evaluate-every 1 --gpu=0 --n-hidden 200 --self-loop --decoder convtranse --encoder uvrgcn --layer-norm --weight 0.5  --entity-prediction --relation-prediction --add-static-graph --angle 10 --discount 1 --task-weight 0.7 --gpu 0 --test

The following command performs multi-step inference and prints the evaluation results (without ground truth history).

python main.py -d ICEWS14s --train-history-len 3 --test-history-len 3 --dilate-len 1 --lr 0.001 --n-layers 2 --evaluate-every 1 --gpu=0 --n-hidden 200 --self-loop --decoder convtranse --encoder uvrgcn --layer-norm --weight 0.5  --entity-prediction --relation-prediction --add-static-graph --angle 10 --discount 1 --task-weight 0.7 --gpu 0 --test --multi-step --topk 0

Change the hyperparameters

To get the optimal result reported in the paper, change the hyperparameters and other experiment set up according to Section 5.1.4 in the paper (https://arxiv.org/abs/2104.10353).

Citation

If you find the resource in this repository helpful, please cite

@article{li2021temporal,
  title={Temporal Knowledge Graph Reasoning Based on Evolutional Representation Learning},
  author={Li, Zixuan and Jin, Xiaolong and Li, Wei and Guan, Saiping and Guo, Jiafeng and Shen, Huawei and Wang, Yuanzhuo and Cheng, Xueqi},
  booktitle={SIGIR},
  year={2021}
}

re-gcn's People

Contributors

Stargazers

Watchers

re-gcn's Issues

OSError: [Errno 22] Invalid argument: '../models/ICEWS14s-uvrgcn-convtranse-ly2-dilate1-his3-weight:0.5-discount:1.0-angle:10-dp0.2|0.2|0.2|0.2-gpu0'

I think the 'best_mrr' can not be updated, could you please tell me where is the problem..It will be a great help to me.

Confusion about baseline

I'm very sorry to be a brother. Can you publish or give me the code of RGCRN in baseline, which will be very helpful to me. If you can, I hope you can send it to my email [email protected].

Filtered Result

Hi,

very nice work!
I am just wondering why you didn't report filtered results in your paper, since filtered results are in most cases much fairer than raw results.

Could you please also post the filtered results of your model?
Thanks!

Zifeng

你好，请问以下下面的错误怎么解决？

Traceback (most recent call last):
File "E:\知识图谱\RE-GCN-master\src\main.py", line 442, in
run_experiment(args)
File "E:\知识图谱\RE-GCN-master\src\main.py", line 234, in run_experiment
loss_e, loss_r, loss_static = model.get_loss(history_glist, output[0], static_graph, use_cuda)
File "E:\知识图谱\RE-GCN-master\src\rrgcn.py", line 212, in get_loss
evolve_embs, static_emb, r_emb, _, _ = self.forward(glist, static_graph, use_cuda)
File "E:\知识图谱\RE-GCN-master\src\rrgcn.py", line 160, in forward
temp_e = self.h[g.r_to_e]
IndexError: tensors used as indices must be long, byte or bool tensors，代码没动过，除了改了一下UTF-8的编码格式，剩下的都是Github上一样的，为什么会报这种错误？

OSError: libcublas.so.10: cannot open shared object file: No such file or directory

Did you meet this problem before?

Ask for codes

Hi, I have read your paper and that is very helpful! I would be appreciated if you commit the codes of this paper:-)

Adaptation to a "different" static graph

Hi!

First of all, thank you for sharing the code for your publication.

I am working to adapt your framework to my problem, where "static" relations regard nodes which are already in the knowledge graph. As a consequence, when setting static_graph.ndata['h'] (rrgcn.py, line 146), I don't need to concatenate dynamic embeddings with word embeddings (which I have deleted), and thus when I take the output of the static layer I don't have to limit to the first self.num_ents elements.

However, with this settings, the entity-prediction loss is nan. I've tried to solve this problem by setting self.h as the sum of the dynamic and static embeddings (rrgcn.py, line 150 and 176). In this way, the problem with the entity-prediction loss is solved, but the static loss has very big values (>2000) and does not decreases over the training epochs. The result is that the static graph doesn't allow to improve performance.

Since this is a methodological question rather than a coding issue, I was wondering if you could help me with your feedback and guidance. Your ideas on how to adapt the methodology and code to work with my task would have a great value to me.

Thanks,
Marco

Confusion about dataset and experiment

Hello, I have some confusions about GEDLT dataset. The training set of GDELT is 1734399 in your paper, but it's down to about a million in the code. May I ask why the data set in the experiment is different from the data set in the paper?

The sceond is when I use CyGNet to experiment on the ICEWS14 dataset in your code, I get better results than your paper.
The result I have gotten is below:

python test.py --dataset ICEWS14s
Namespace(alpha=0.5, batch_size=1024, counts=4, dataset='ICEWS14s', entity='object', gpu=0, hidden_dim=200, lr=0.001, n_epochs=30, raw=False, regularization=0.01, time_stamp=1, valid_epoch=5)
num_times 366
start object testing:
Using best epoch: 7
test object-- Epoch 0007 | Best MRR 0.5029 | Hits@1 0.4366 | Hits@3 0.5419 | Hits@10 0.6185
start subject testing:
Using best epoch: 8
test subject-- Epoch 0008 | Best MRR 0.4795 | Hits@1 0.4084 | Hits@3 0.5197 | Hits@10 0.6051

final test --| Best MRR 0.4912 | Hits@1 0.4225 | Hits@3 0.5308 | Hits@10 0.6118

关于论文中报告的数据集信息和代码公开的数据集统计信息不一致

Hello!

It seems like there might be a discrepancy between the reported number of entity categories in the paper (6,869) and the stat file in the publicly available data-release.tar.gz, which mentions 7,128 entity categories. This could be due to a typo or a difference in the dataset versions.

Thanks for your answers!!!

Confusion about the experimental setup

I have read your paper and that is very helpful! I'm curious about the experimental setup without ground truth. I tried some experiments, but I didn't get the result that you wrote in the paper without ground truth. It would be very helpful if you could give the experimental setup for this part.

cuda out of memory

I tried to train a model on my own dataset, but I came with the problem "cuda out of memory", is there any good way to solve that problem?

The results for WIKI dataset are different from the paper

In your paper, the Hit@10 for WIKI is 53.88 or 53.88. But I run the code, and get 0.833 Hit@10, which is almost the same as the Yago dataset.
I am curious how you do the experiments. Can the model distinguish between Yago and WIKI?

confusion about parameter(skip_connect)

i found skip_connect=True never works in your code.
it is used in a layer as follows:

            for i, layer in enumerate(self.layers):
                layer(g, [], init_rel_emb[i])

and layer calls for the following code,and bool expression len(prev_h) != 0 and self.skip_connect is forever False.

    def forward(self, g, prev_h, emb_rel):
        self.rel_emb = emb_rel
        if len(prev_h) != 0 and self.skip_connect:
            skip_weight = F.sigmoid(torch.mm(prev_h, self.skip_connect_weight) + self.skip_connect_bias)

so this parameter doesn't work,is there something wrong?

confusion about dataset

can someone tell me what's "24" and "0" mean?What i know is all lines with 24 are events happens on 2018-01-02,but i don't know why 2018-01-02 is encoded as 24 and what does the last column represent?

command line for running on YAGO and wiki

Can you share the command lines for running on YAGO and wiki?

I got such an error

FileNotFoundError: [Errno 2] No such file or directory: './data/YAGO/e-w-graph.txt'