Join GitHub today
GitHub is home to over 40 million developers working together to host and review code, manage projects, and build software together.Sign up
[Bugfix] Correct the loss function in RGCN model #1217
As mentioned in #1204 , the loss function in our RGCN implementation is problematic. This PR fix the issue.
Please feel free to remove inapplicable items for your PR.
I already fixed it in tensorflow at https://github.com/dmlc/dgl/blob/master/examples/tensorflow/rgcn/entity_classify.py#L109
@jermainewang Nope. I get the following mean accuracies over 10 repeats (standard deviation in brackets):
aifb: 0.961 (0.0194)
BGS is a little below what we had in the original paper, but the variance is high, so I think we just got lucky. All hyperparameters are the same as in the paper, and the weight initialization is taken from your code (ie. uniform xavier with relu gain everywhere).
I strongly recommend averaging over 10 repeats, since the test sets are so small, but none of my runs on BGS are as low as .75.