- Download the
notebook
. You can upload to Google Colab or run on your own computer. - Add files in ‘data’ to your working directory.
- Run cells in 1 to 4. Skip 5.
- To see sample plot, run cells in 6
- Download the
notebook
. You can upload to Google Colab or run on your own computer. - Add files in ‘data’ to your working directory.
- Run cells in 1 to 3, and data loading cells in 4-1 and 4-2
- Run cells in 5
Below is a brief description of the notebook.
- load_ZINCdataset1 and load_CEPdataset1 returns data including one-hot encoded matrix of atomic features
- load_ZINCdataset2 and load_CEPdataset2 returns data including atomic feature and molecular feature encoded matrix
- implementation of skip connection and gated skip connection
implementations of Attention and gate-augmented graph convoutional networks
- Graph Convolutional Networks (GCN) with skip connection
- Graph Convolutional Networks (GCN) with gated skip connection
- Graph Attention Networks (GAT) with skip connection
- Graph Attention Networks (GAT) with gated skip connection
Graph Isomorphic Networks
- Graph Isomorphic Networks (GIN)
- Graph Convolution Netowkrs (GCN)
- Graph SAmple and aggreGatE (SAGE); mean aggregator version
3 Modified SAGE layer
- run torch models GCNwithSkip, GCNwithGate, GATwithSkip, GATwithGate, and torch_geometric models GIN, GCN, SAGE, modified_SAGE with data loaded by 1-1 functions
- run the same models by loading data (1-2) with args.mol_dim = 0
- run the same models by loading data (1-2) with args.mol_dim = 19
- hyperparameter settings for best performance in LogP, TPSA, SAS, CEP
- the pretrained model load function
- Reproducing result
- Compare three feature encodings
- saved all experiment records (288 model configuration and 3 type data representation method)