/CCN

Code for paper "Multi-label Classification Neural Networks with Hard Logical Constraints"

Primary LanguagePython

CCN

Code and data for the paper "Multi-Label Classification Neural Networks with Hard Logical Constraints"

Train CCN

In order to train the network use the file train.py. An example on how to run it for the dataset emotions is the following:

python train.py --dataset "emotions" --num_classes 6 --seed "$seed" --split "$split" --device "$device" --lr "$lr" --dropout "$dropout" --hidden_dim "$hidden_dim"  --num_layers "$num_layers" --weight_decay "$weight_decay" --non_lin "$non_lin" --batch_size "$batch_size"

train.py saves a pickle file for each execution in the hyp folder. Each pickle file stores the value of the hyperparameters and of the validation loss.

Test CCN

In order to test the network use the file test.py. An example on how to run it for the dataset emotions is the following:

python test.py --dataset emotions --seed "$seed" --device 1&

test.py finds the combinations of hyperparameters that results in the best validation loss, train the relative model, and then writes the results in the results folder.

Reference

@article{giunchiglia2021,
  author    = {Eleonora Giunchiglia and Thomas Lukasiewicz},
  title     = {Multi-Label Classification Neural Networks with Hard Logical Constraints},
  journal   = {Journal of Artificial Iintelligence Research (JAIR)},
  volume    = {72},
  year      = {2021}
}