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

Citeseer data set accuracy #14

Closed xuhaiyun42 closed 5 years ago

xuhaiyun42 commented 5 years ago

Hello!I used the GAT network to run the citeseer database, but the accuracy could not reach 72.5, only 70.3. How did you set the parameters to run so high?

PetarV- commented 5 years ago

Hello,

Thanks for your issue!

Regarding the Citeseer setup, we have found that early stopping just on the accuracy (rather than loss and accuracy) yielded better results.

Here is a relevant code segment:

if val_acc_avg/vl_step >= vacc_mx:
    if val_acc_avg/vl_step >= vacc_mx:
        vacc_early_model = val_acc_avg/vl_step
        vlss_early_model = val_loss_avg/vl_step
        saver.save(sess, checkpt_file)
    vacc_mx = np.max((val_acc_avg/vl_step, vacc_mx))
    vlss_mn = np.min((val_loss_avg/vl_step, vlss_mn))
    curr_step = 0

Hope that helps! Note that the standard deviation on Citeseer is large (0.7) so it might take multiple runs to achieve a satisfactory accuracy. For example, I had five runs in a row with 73.1%, 74.2%, 71.9%, 73.1%, 70.9% under this configuration.

Thanks, Petar

xuhaiyun42 commented 5 years ago

Thank you very much for the reply.

Thanks, Xu Haiyun