This is the official code of paper Ambiguity-Aware Abductive Learning, In: Proceedings of 41st International Conference on Machine Learning.

Set up

# Require swi-prolog >= 8.0
# In ubuntu system  22.04 or higher
sudo apt install swi-prolog
# Otherwise, please download the specific version of swi-prolog and install it manually.

# Create Conda Enviroment
conda create --name abl python=3.10 
conda activate abl 

# Login your wandb account(if not, the logging process will encounter error.)
# The main results can be seen on the **wandb webpage, check it please!**
wandb login

Experiment reproduce

Digit Addition

cd examples/addition

# dataset in [MNIST, KMNIST, SVHN, CIFAR]
# digit_size in [1, 2, 3, 4]
# Note: if digit_size is 2,3,4 the learning process will not be very quick, be patient.
python wsabl.py --dataset $dataset --digit_size $digit_size

Handwritten Formula Recognition

# HWF
cd examples/hwf 
cd datasets
tar xf data.tgz 
cd ..
python wsabl.py 

# HWF-CIFAR
cd examples/hwf-cifar
cd datasets
tar xf data.tgz 
cd ..
python wsabl.py

# HWF-SVHN
cd examples/hwf-svhn
cd datasets
tar xf data.tgz 
cd ..
python wsabl.py

AcKnowledgement

Thanks for the great libs:

• ABLKit: https://github.com/AbductiveLearning/ABLkit (The code is mainly based on a previous commit of ABLKit, although there should be a litter different between APIs and details.)
• Weights & Bias: https://wandb.ai/site (for logging)

Others

The name wsabl is a short for Weakly Supervised Abductive Learning, which is a previous and initial naming way of A3BL. Any question or suggestion, please contact me: [email protected](prefered).

Licecense

This project is licensed under the terms of the MIT license.