This repository contains the implementation for ARGA as described in the AAAI-23 paper Graphs, Constraints, and Search for the Abstraction and Reasoning Corpus. Some code were edited for formalism, which caused some small variations in the results as presented in the paper.
To run ARGA on any ARC task, simply call
python main.py taskid.json tasktype(training/evaluation)
for example, to run the task #ddf7fa4f:
python main.py ddf7fa4f.json training
the results will be stored under the solutions folder and the visualization for the results will be stored under the images folder.
Example tasks used for illustration in the paper are as follows:
- recolor task: d2abd087.json
- dynamic recolor task: ddf7fa4f.json
- movement task: 3906de3d.json
- augmentation task: d43fd935.json
The ids for the 160 selected tasks can be found under dataset/subset.
The full set of the ARC dataset created by Chollet can be found here as well as under the dataset folder.