irl-text-generation

Joint work with Cailin Winston and Peter Michael. This repo contains code to run inverse reinforcement learning (IRL) applied to the problem of text generation. The full write-up is in irl_text_generation.pdf. This project was the response to an open-ended final project in the Generative Models course at the University of Washington. The project idea is based on the paper 'Toward Diverse Text Generation with Inverse Reinforcement Learning' https://arxiv.org/abs/1804.11258 IJCAI2018 with some additional modifications.

• Overview
• System Requirements

Overview

To reproduce the results, you run the imagecoco/train.ipynb notebook within a GPU-accelerated Google Colab enviornment, filling in the TODOs that specify save folders.

System Requirements

Hardware requirements

The code requires a CUDA-enabled GPU.

Python Dependencies

The code mainly depends on the Python machine learning stack. Requirements can be found in environment.yml, and can be installed via conda.

conda create -n rl-project python=3.6
conda env update -n rl-project -f environment.yml
conda activate rl-project