This is the repo for our TMLR paper "Identifying Axiomatic Mathematical Transformation Steps using Tree-Structured Pointer Networks" (https://openreview.net/forum?id=gLQ801ewwp).
In this Readme we explain how to generate data, train models and evaluate them.
For generating data and inference install the dependencies in requirements.txt.
The code of the dataset generator can be found in the folder generator.
Run the code as follows: python3 eqGen.py $n_equiv $n_trans $axioms_path where $n_equiv is the number of start equations, $n_trans is the number of axiomatic transformation steps and $out_path is the folder where the generated data should be saved, e.g. python3 eqGen.py 500 5000 /tmp/data.
We generally set n_eqiv=10000, n_trans=10*n_equiv.
You have to copy the file axioms.csv (or the file which specifies your own axioms) to $out_path before running the code.
It is advisable to run multiple instances of the generator in parallel and merge the data afterwards.
The usage of TreePointerNet is described here: https://github.com/sj-w/tree_pointer_net
The baseline models are just regular fairseq models. See https://github.com/facebookresearch/fairseq on how to use this toolkit.
The code for evaluation is located in inference. Run the script run_eval_steps_cluster.sh $checkpoints_folder $data $mode where $checkpoints_folder is the folder to the fairseq checkpoints of the trained models, $data is the folder containing your test data and $mode is either "tree" (for TreePointerNet) or "sequence" (e.g. transformer).
If you find our code helpful, please cite the following paper:
Coming soon.