object reference synthesis

Generating Programmatic Referring Expressions via Program Synthesis

Install

Setup the environment

In this project, we used pytorch-geometric and the dependancies of it.

Method 1. Import the environment to your anaconda (Recommended) In this directory, there is a torch_geometric.yml file. You can import the environment to your anaconda to install all the dependancies. https://docs.conda.io/projects/conda/en/latest/user-guide/tasks/manage-environments.html

Method 2. Use the dockerfile In this directory, there is a Dockerfile. You can build an docker image here and run the result inside.

docker build .
docker run -it <DOCKER IMAGE NAME>

then in the container, you need to activate the conda environment.

conda init bash 
exec bash 
conda activate pytorch_geometric

Method 3. build from scratch You can install pytorch geometric according to the documentation here https://pytorch-geometric.readthedocs.io/en/latest/notes/installation.html

Setup the directories

Type make in the current directory to setup the directory hierachies.

make 

Run the scripts

In this directory, there is a run.py which invokes different scripts to execute different tasks. You can tune the parameters in the run.py to change the settings. Further, there are more parameters you can change in the src/cmd_args, and src_prob/cmd_args.

Note: Please remember that, if there is a previous model on the model path, we will load the model and continously train on it. So if you want to retrain the model using another method, move the old model to some other directory.