atom_init.json files for pre-trained models

[boriskiefer]

Good Evening:

I am tryng to run CGCNN2 on the pre-trained bandgap model. I followed the formation_energy example, converted jupyter notebook to run on linux CDL, pointed to the model, for simplicity I used the 10 cif files, in that directory. Run completed and I inspected my saved bandgap file, suggesting nonsense output. For verification I downloaded mp-66 and mp-985 and ran in separate directory, also nonsense. I think the root cause of the problem may that cgcnn2 expecs an atom_init.json file that matches the pretrained model weights. I cannot find the correspodning pair on github repo or in
my pretrained model subdirectory of cgcnn2. Hence my question, pre-trained model specific atom_init.json files be made available? This would be very helpful in running the pretrained models. I apologize is my understading of the CGCNN2 machinery is incorrect in whcih case any advise on my error would be appreciated.

Thanks -- Boris Kiefer

[jcwang587]

Hi @boriskiefer, thank you for trying this package and raising the issue. The pretrained models and atom_init.json in cgcnn2 are a fork from the original cgcnn repository. I believe this atom_init.json was the one used for the pretrained models in the original paper.

There are some differences in the cgcnn2 framework compared with the original, but they are mainly to improve speed and suppress deprecation warnings. I haven’t yet tested whether these changes would affect the pretrained models’ performance. I’m sorry I don’t have a quick fix for this problem, but I will add this to my to do task retraining the CGCNN models on the latest Materials Project data. Be aware that, although CGCNN could be easy to run, it may not be a state-of-the-art model for a universal property prediction.