Code for the paper "Learning with Logical Constraints but without Shortcut Satisfaction"
numpy
pytorch
For each task, run the following command.
python logic_training.py --num_labeled 100 --constraint True --trun True --exp_name logic_trun \
--net_type densenet100 --adam_lr 0.01 --constraint_weight 1.0
where the parameter "trun" is set to True to enable the truncation of Gaussian distribution.
To reproduce the experimental results, Run the command sh run.sh
-
We also provide baseline methods in MNIST task, one can refer to
/mnist_exp/baselinesfor more details. -
Our another work ("Softened Symbol Grounding for Neuro-symbolic Systems") can also avoid the shortcuts, but do not need to additionally define the dual variable.