kaist-amsg/cgcnn-hd

CGCNN-HD (https://dx.doi.org/10.1021/acs.jcim.0c00003)

The CGCNN-HD is a python code for dropout-based uncertainty quantification for stability prediction with CGCNN (by T. Xie et al.) developed by prof. Yousung Jung group at KAIST (contact: [email protected]).

Developers

Juhwan Noh

Dependencies

• Python3.6
• Numpy
• Pytorch == 0.4.1.post2 (CUDA 8.0)
• Pymatgen
• Sklearn
• ASE

How to use

1. Database setting Reference formation energy value of the previous our ChemComm paper (Chem. Commun.,2019,55,13418-13421) can be found in Mg-Mn-O_database/MgMnO_form_e.data.k500.json

If you want to use org_cif database

• cd Mg-Mn-O_database
• tar xvf org_cifs.tar
• cp id_prop.r4_nn8.orgcif.csv org_cifs/id_prop.csv
• cp atom_init.json org_cifs/

If you want to use scaled database

• cd Mg-Mn-O_database
• tar xvf lattice_scaled.tar
• cp id_prop.r4_nn8.scaled.csv lattice_scaled/id_prop.csv
• cp atom_init.json lattice_scaled/

2. Dropout Sampling

If you want to use org_cif database

• Currently, crystal graph is constructed only if maximum number of neighboring atom = 8 and cutoff radius = 4A
• Change root_dir = '/your/data/path/' in dropout_sampling.py to Mg-Mn-O_database/org_cifs/
• python dropout_sampling.py cgcnn_hd_rcut4_nn8.best.pth.tar
• You may get dropout_test.csv file (name,predicted mean,predicted standard deviation)

If you want to use scaled database

• Currently, crystal graph is constructed only if maximum number of neighboring atom = 8 and cutoff radius = 4A
• Change root_dir = '/your/data/path/' in dropout_sampling.py to Mg-Mn-O_database/lattice_scaled/
• python dropout_sampling.py cgcnn_hd_rcut4_nn8.best.pth.tar
• You may get dropout_test.csv file (name,predicted mean,predicted standard deviation)

3. Training with your own database

• Currently, crystal graph is constructed only if maximum number of neighboring atom = 8 and cutoff radius = 4A
• Change root_dir = '/your/data/path/' in model_train.py to path for your own dataset
• Set MYPYTHON="your/python/path" and MODELPREF="your/model/pref" in training.sh with your own setting
• sh training.sh

cgcnn-hd's Issues

Update dependencies

Just a head's up, there are two clarifications that should be made to the list of dependencies:

1. ASE is not mentioned
2. Python 3 should be Python 3.6. Anything >= 3.7 causes an issue with pytorch's async flag.

Other than that, I was able to run the training.

Undefined variable dropout_samples in cgcnn_hd.py

In cgcnn_hd.py, the sampling function returns dropout_samples, which is never defined. There also appears to be several other undefined variables. It seems dropout_sampling.py is not up-to-date.