BOML

Bayesian optimized machine learning. Under continual development with addition of new models and architectures.

Until further publication, please cite: Maffettone P.M., Cooper, A.I. Deep learning from crystallographic representations of periodic systems. In: Fall 2019 ACS National Meeting; 2019 Aug 25--29; San Diego; ACS; 2019. No CINF 84.

Abstract here.

Adding a new model requires submodule in ml that contains: models.py {change appopriate conditional statement in training.py} parameters.py {load_metaparameters, gen_hyperparameters}

Adjustments must be made to: boml.optimization.py {Optimizer.fetch_model_functions}

Adjustments can be made to: boml.utils.sanity_checks.py boml.utils.defaults.py

Classification

Classification models should have X values in sub-directories that correspond to their class. The parent directory is used as input for the training.

Regression

Regression models should have all X values in a single directory. There should be a corresponding DataFrame pickle or csv file, with a column header of "Name" and column header of "XXX".
"Name": X value file name WITHOUT extension "XXX": Regression target, defaults to Energy

Requires updates on addition of new models

boml.optmization.Optimizer.fetch_model_functions()
boml.ml.XXX_models.training.gen_model()