search

PetarV- / GAT

Graph Attention Networks (https://arxiv.org/abs/1710.10903)
https://petar-v.com/GAT/
MIT License
3.18k stars 642 forks source link

Pytorch implementation #3

Closed Diego999 closed 6 years ago

Diego999 commented 6 years ago

Hi,

I was really interested to use your model but unfortunately it is in Tensorflow (or Keras for the unofficial implementation). I propose here the Pytorch version (https://github.com/Diego999/pyGAT) where I obtained between 83.6-84.6 for the accuracy on the transductive learning in the cora task. If you would like to add it in your readme (didn't want to do a pull request for this).

Best,

PetarV- commented 6 years ago

Hi Diego,

Thank you for your interest in GAT! Especially for putting in the effort to reproduce our result in PyTorch. This should be very helpful indeed.

Would you mind adding the citation reference to your README? I'll then gladly link to your repository from ours.

Thanks, Petar

Diego999 commented 6 years ago

Hi Petar,

Yes definitely ! It's done.

Best,

JsonAC commented 5 years ago

Hi,

I was really interested to use your model but unfortunately it is in Tensorflow (or Keras for the unofficial implementation). I propose here the Pytorch version (https://github.com/Diego999/pyGAT) where I obtained between 83.6-84.6 for the accuracy on the transductive learning in the cora task. If you would like to add it in your readme (didn't want to do a pull request for this).

Best,

hi, the training/testing split in your Pytorch implementation is different from that in the paper(GCNs, GAT). When I change your utils to follow the original split and change the random seed 10 times, the average accuracy falls to about 82.6%

KL-ice commented 4 years ago

Hi, I was really interested to use your model but unfortunately it is in Tensorflow (or Keras for the unofficial implementation). I propose here the Pytorch version (https://github.com/Diego999/pyGAT) where I obtained between 83.6-84.6 for the accuracy on the transductive learning in the cora task. If you would like to add it in your readme (didn't want to do a pull request for this). Best,

hi, the training/testing split in your Pytorch implementation is different from that in the paper(GCNs, GAT). When I change your utils to follow the original split and change the random seed 10 times, the average accuracy falls to about 82.6%

Hello, I have encountered the same problem as you. Have you solved the problem that the random seed has a big influence on the code of the pytorch version? Thanks， Ice.