CGEM: Concept Graph Embedding Models

This repository contains the code for our concept graph embedding models (CGEM). The codebase is implemented using PyTorch and tested on Ubuntu 16.04.4 LTS

Prerequisite

Configure environment

Install Anaconda.

Create and activate a virtual environment.

conda create --name cgem python=3.9
conda activate cgem

The code is tested with python 3.9, cuda == 10.2, and pytorch == 1.12.1. Install the required additional packages.

pip install -r requirements.txt

Download dataset

All datset must be downloaded to a directory '../_dataset' and must follow the below organization.

├──_dataset/
    ├──cub/
        ├──images/
        ...
├──ConceptGraphEmbedding/
    ├──train.py
    ├──config.py
    ├──dataset/
    ├──models/

We refer to CEM's repository to download and prepare data.

Directories

• dataset: contains scripts to load the datasets used in our experiments
• models: contains the implementations of our concept encoder and concept GCN models