Below you can find a outline of how to run my solution to Kaggle's Abstraction and Reasoning Challenge.

Tested computer setup

• Ubuntu 18.04.4 LTS
• Intel® Core™ i7-7700HQ CPU @ 2.80GHz × 8
• 16GB RAM
• Python 3.6.9
• g++ 7.5.0

Any Python 3 and g++ supporting c++17 should work.

Running on public data

The comptition data is already in the "dataset" folder for conveniece.

You can run the model on the evaluation dataset using depth 2 with (takes 70 seconds on my computer):

python3 run.py

You can view a summary of the results with:

python3 summary.py

It should give 129 / 419 correct predictions.

To run using depth 3, change "run_depth" to 3 in run.py
To run on the training dataset, change "sample_dir" to "training" on line 79 in src/runner.cpp, and set "inds = range(0,416)" in summary.py

Running on test data

To run the full model and produce precictions on the test set (takes 9 hours), change "eval" to 1 on line 75 on src/runner.cpp and run

python3 safe_run.py

This produces the output file named "submission_part.csv", which can be renamed to "submission.csv" to submit to the competition.

Citation

If you find this repository helpful, please cite:

@software{wind2020dsl,
  title  = {DSL solution to the ARC challenge},
  author = {Wind, Johan S.},
  url    = {https://github.com/top-quarks/ARC-solution},
  year   = {2020}
}