jinheonbaek/rgcn

Pytorch implementation of a RGCN Link Prediction Model

PyTorch RGCN (Link Prediction)

PyTorch implementation of Relational Link Prediction of RGCN (Modeling Relational Data with Graph Convolutional Networks). The code is sparsely optimized with torch_geometric library, which is builded based on PyTorch.

About

This is the original tensorflow implementation of link prediction of RGCN: https://github.com/MichSchli/RelationPrediction

Also, most of the functions in utils.py are brought from the following repository: https://github.com/dmlc/dgl/tree/master/examples/pytorch/rgcn, where this repositroy is based on another popular graph library, namely dgl.

Requirements

This code is lastly tested with:

• python 3.7.x
• pytorch 1.7.x
• torch_geometric 1.7.x, with torch_scatter 2.0.6 and torch_sparse 0.6.9

rgcn's Issues

So how did you finally solve this problem

How is a MRR score of 0.20 possible?

I trained with your code and got an MRR of 0.2. In the original paper they only reach an MRR score of 0.156. How is my score with a model with fewer parameters than the original approach possible?

Which version do u use ? I have some troubles in torch_sparse

error

Namespace(dropout=0.2, evaluate_every=500, gpu=-1, grad_norm=1.0, graph_batch_size=30000, graph_split_size=0.5, lr=0.01, n_bases=4, n_epochs=10000, negative_sample=1, regularization=0.01)
load data from ./data/FB15k-237
num_entity: 14541
num_relation: 237
num_train_triples: 272115
num_valid_triples: 17535
num_test_triples: 20466
Traceback (most recent call last):
  File "main.py", line 121, in <module>
    main(args)
  File "main.py", line 48, in main
    test_graph = build_test_graph(len(entity2id), len(relation2id), train_triplets)
  File "utils.py", line 156, in build_test_graph
    data = Data(edge_index = edge_index)
  File "C:\Users\xxxxxxxxx\.conda\envs\a4\lib\site-packages\torch_geometric\data\data.py", line 88, in __init__
    raise ValueError(
ValueError: Argument `edge_index` needs to be of type `torch.long` but found type `torch.int32`

Which version do u use ? I have some troubles in torch_scatter

It is a great jobs and I'm very appreciate about it.
But when i trying to : pip install torch_scatter the terminal returns a stack red error seems that my cuda version not fitted with it
So could u please tell me about the version u use ?
Thanks a lot~

work on entity alignment

does r-gcn works on entity alignment (existing R-GCN seems only focus on classification and link prediction)?

If yes, May I know do you have any working code on this?

or does this link prediction model works directly on that? Since “Links between two nodes exist” is very similar to “two nodes referring to same real-world object”

Why did it appear killed when training the model

best_mrr_model.pth

__where can I find the 'best_mrr_model.pth'?

Why building test graph requires train_triplets?

Hi Dear Author, in this line, building the test_graph has train triplets as input, isn't it test triplets actually?

RGCN model conv1 is applied two times.

In the RGCN model, conv1 is applied twice in succession. Is that intended?

x = self.conv1(x, edge_index, edge_type, edge_norm)        
x = F.relu(self.conv1(x, edge_index, edge_type, edge_norm))

Can't reproduce results in the paper

Hi,
Thanks for your excellent job! This is the only repo I can find that is implemented in torch_geometric. But I can't reproduce results using your code. Can you show me your results and configurations.
Thanks.

The memory overflow

I use your code as a part in our experiment.
But there is a problem about the test_graph.
Because the code uses all train triplets to build graph during valid and test, so when I test the model, 64G memory is not enough.
So I wish you to give me some suggestions.
Thank you very much!