Learning Decision Trees as Amortized Structure Inference

This repository contains code for Learning decision trees as amortized structure inference.

It is largely based on https://github.com/alexhernandezgarcia/gflownet.

Illustration

Initial setup

For the inital setup, simply follow these steps:

git clone https://github.com/GFNOrg/dt-gfn.git
cd gfn
pip install -r requirements.txt

To setup the cython files, please make sure to run:

cd gflownet
python setup.py build_ext --inplace

Quickstart

You can train a GFlowNet to sample a single decision tree on the Iris dataset with the following command:

python main.py +experiments=tree_acc

To make sure everything works seamlessly, you might want to use the following command for debugging:

HYDRA_FULL_ERROR=1 python main.py +experiments=tree_acc

If tree plotting with Graphiz is an issue, either make sure to install it or use the following code:

HYDRA_FULL_ERROR=1 python main.py +experiments=tree_acc ++env.test_args.top_k_trees=0