The reference codebase for the paper "Graph Structured Prediction Energy Networks," appeaing in NeurIPS 2019.

This code was developed/run using the following versions of the following libraries; in some cases, newer versions may be acceptable. The primary exception is PyTorch - some of the syntax used here changed in newer versions.

• PyTorch 0.4.1
• torchvision 0.2.0
• numpy 1.15.4
• scikit-image 0.13.1
• tensorflow 1.12.0 (This is for training visualization purposes)
• liac-arff (for reading bibtex)
• torchfile 0.1.0 (for reading bookmarks data files)

Additionally, this code contains a module written in C++; thus, a C++ compiler needs to be installed as well. To make use of LP/ILP inference, you will need a valid Gurobi installation; make sure the GUROBI_HOME environment variable is set to the root dir of this installation.

Make sure to install the library using pip before running (this compiles the C++ code) by running the following command from the root directory:

pip install ./

If you plan on making your own changes, make sure to include the -e flag. Run all scripts from the root directory.

The (compressed) synthetic words datasets are included in the data/ directory. The bibtex/bookmarks datasets can be downloaded here. The script data/arff2tensor.py converts the raw data files into the form used by the training code; see scripts/run_bibtex.sh to see how to run this script. The bookmarks experiment script uses the data files provided by following the instructions listed here. Tagging data can be found here.