alphaILP: Thinking Visual Scenes as Differentiable Logic Programs

Documentation is here!

This is the implementation of the paper:

alphaILP: Thinking Visual Scenes as Differentiable Logic Programs

to appear in Machine Learning, and presented at IJCLR 2022

Hikaru Shindo, Viktor Pfanschilling, Devendra Singh Dhami, and Kristian Kersting.

Installation

The packages are specified in requirements.txt. Please install the packages by:

pip install -r requirements.txt

Training

Kandinsky Patterns Twopairs Dataset:

python src/train.py --dataset-type kandinsky --dataset twopairs --batch-size 1 --no-cuda --n-beam 5 --t-beam 5

CLEVR-Hans3 Dataset:

python src/train.py --dataset-type clevr --dataset clevr-hans0 --batch-size 1 --no-cuda --n-beam 15 --t-beam 5
python src/train.py --dataset-type clevr --dataset clevr-hans1 --batch-size 1 --no-cuda --n-beam 15 --t-beam 6
python src/train.py --dataset-type clevr --dataset clevr-hans2 --batch-size 1 --no-cuda --n-beam 15 --t-beam 7

Pretraining Neural Predicates

Pretraining of neural predicates can be done by, e.g., :

python src/train_np.py --dataset closeby_pretrain --no-cuda --batch-size 2 --lr 1e-1 --epochs 10
python src/train_np.py --dataset online_pretrain --no-cuda --batch-size 2 --lr 1e-1 --epochs 10

LICENSE

See LICENSE. The src/yolov5 folder is following GPL3 license.