materialsproject
add megnet and SchNet
previous results got corrupted, need to redo the experiment
also should add SchNet, as others are interested in it
I can add MegNet and SchNet results on the benchmarks using kgcnn. Would that work for MatBench? I opened a pull request with first training result.
Yes that would be great! Especially if you are planning on submitting for multiple tasks! Have you already run the training and benchmarking or is this prospective?
Thanks, Alex
On Sep 13, 2022 at 4:08 AM -0700, Patrick Reiser @.***>, wrote:
I can add MegNet and SchNet results on the benchmarks using kgcnn. Would that work for MatBench? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>
Yes, I have run SchNet on all structure benchmarks but wihout hyper-optimization. I created pull request #184 I will run MegNet and DimeNet++ next.
I added MegNet benchmark in pull request #187 . Quick question: If I were to rerun the training with optimized hyperparameter but same model/version, should I overwrite the existing folder or add completely new benchmark?
@PatReis I would suggest submitting an entirely new benchmark, but make it very clear in the long description why a new benchmark was submitted and what the difference is between this benchmark and the previous version.
Some other suggestions: Also make it very clear in writing and in dict form what the hyperparameters are and by what kind of method they were determined. If it was by a method outside of what was already published with MEGNet, make sure to cite that. Adding in what kind of hardware it was trained on might be of use as well (and/or just reporting the time taken for training on each of the training folds as hyperparameter info when you execute .record)
Also, thank you so much for the work on this! It is much appreciated!