Can pytorch version 3dcnn reproduce the results of paper?

[zhuhui-in]

When I tried the 3dcnn pytorch version, can't reproduce the perasonr mentioned in the paper. The training performance on refine_minus_core set is 0.95(pearsonr), the test performance on core set is 0.59(pearsonr). Any suggestions?

[wderekjones]

Hi @zhuhui-in, sorry to hear you are having issues. The original version of the code was written in Tensorflow for the 3D-CNN, so its definitely possible that there will be some differences between that package. and pytorch in terms of their backends. The pytorch version was not developed until after the work was complete, so I can't guarantee that it will behave the exact same. With that said, I would try to enumerate a number of random seeds for training and compute a distribution of values on your test sets to see how that looks. It also seems that you're overfitting in this case, so tweaking some of the regularization parameters might be a good idea too.

If you continue to have issues, please let us know.