`alinemol.hyper`¶

Hyperparameter search spaces for the GNN model families, used with hyperopt-style tuning in the training scripts.

from alinemol.hyper import init_hyper_space

space = init_hyper_space("GCN")   # search space for the GCN model

Building a search space¶

init_hyper_space(model)

Initialize the hyperparameter search space

model : str Model for searching hyperparameters

dict Mapping hyperparameter names to the associated search spaces

alinemol.hyper.hyper also defines per-model hyperparameter dictionaries that init_hyper_space selects from:

Object	Model
`common_hyperparameters`	Shared across models (learning rate, dropout, …)
`gcn_hyperparameters`	Graph Convolutional Network
`gat_hyperparameters`	Graph Attention Network
`weave_hyperparameters`	Weave
`mpnn_hyperparameters`	Message Passing Neural Network
`attentivefp_hyperparameters`	AttentiveFP
`gin_pretrained_hyperparameters`	Pretrained GIN
`nf_hyperparameters`	Neural Fingerprint

See scripts/clf_train_gnn.py for how these feed into the training loop.