Examples

Model parameters

Optimal model parameters for the training of the models implemented in the toolbox are stored in model_params. The tuning has been performed on a POWER8 cluster, hence the models might be too large for training on a laptop. The parameter files are in json format and each corresponds to a different encoder:

• Contextual attention encoder (ca.json)
• Dense encoder (dnn.json)
• Multiscale convolutional attention encoder (mca.json)
• Recurrent encoder (rnn.json)
• Sequence attention encoder (sa.json)
• Convolutional encoder (scnn.json)

Toy data

In data/train and data/test we include a small collection of TFRecords compatible with the format used by the toolbox.

Run a example

After installation, to run a training on the toy examples use training_paccmann passing the data, the desired encoder and the corresponding model parameters. Assuming you are inside this folder ("examples"), run:

training_paccmann data/train data/test /path/to/store/model/ mca model_params/mca.json smiles_atom_tokens,selected_genes_20

For details on the TFRecords schema check paccmann/datasets.py.

Run example (baseline models)

You can compare the performance of the different NN architectures with off-the-shelf regression models (SVR, RF, AdaBoost etc.) by running:

training_baseline data/train data/test /path/to/store/model/

Please note that performance scores are not representative due to the size of the dataset.