Overview:

Folders:

data - Contains all the dataset used in this paper. models/box/gumbel_bce_box.py - Contains the implementation of Tasks for GumbelBox. models/box/bce_box.py - Contains the implementation of Tasks for SmoothBox and Gaussian.

configs - Contains ".jsonnet" files for each experiemnt that we report. These files contains the best hyperparameter for the corresponding task that file refers to.

Submodules:

Box embedding implementation - pip install git+https://github.com/ssdasgupta/boxes.git --no-deps

Dataset pipeline implementation - pip install git+https://github.com/ssdasgupta/datasets.git --no-deps

Install:

conda create -n neurips2020 python=3.6
conda activate neurips2020
pip install -r requirement.txt
pip install git+https://github.com/ssdasgupta/datasets.git --no-deps
pip install git+https://github.com/ssdasgupta/boxes.git --no-deps

Run:

Note that config files corresponding to all the experiments are in "configs" folder. They are in ".jsonnet" format.

Run the following:

allennlp train configs/choose_you_config_file -s logs/Exp_you_want_to_run/ --include-package models --include-package datasets  -f

Configs

For experiments (Section 5.3) with WordNet's full noun hierarchy the config files will be similar to the following: wordnet_full_{model_name}_{with or without(wo)}_reg.jsonnet

For experiments (Section 5.1) on ranking task on different tree-structures: {tree_name}{model_name}{number_of_dim}.jsonnet

With this information, the config names are self explanatory.

Results

The value of different metrics for the experiment can be found in the log_directory/metrics.json.