Skip to content
EGG: Emergence of lanGuage in Games
Jupyter Notebook Python
Branch: master
Clone or download
mnoukhov and eugene-kharitonov fix for torch 1.2 (#28)
fix for torchvision 0.4 which requires torch 1.2
Latest commit 60ed159 Aug 13, 2019
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
.circleci Initial commit Jun 6, 2019
docs Update nest.md Jun 6, 2019
egg checkpointing under slurm (#27) Aug 9, 2019
tests Fix tests (#25) Aug 6, 2019
tutorials Move remaining trainer functionality to the new callback API (#17) Aug 1, 2019
.gitignore Initial commit Jun 6, 2019
CHANGELOG Initial commit Jun 6, 2019
CHANGELOG.md Initial commit Jun 6, 2019
CODE_OF_CONDUCT.md Initial commit Jun 6, 2019
CONTRIBUTING.md Initial commit Jun 6, 2019
LICENSE Initial commit Jun 6, 2019
README.md Update README.md Aug 6, 2019
pytest.ini Initial commit Jun 6, 2019
requirements.txt fix for torch 1.2 (#28) Aug 13, 2019
setup.py Initial commit Jun 6, 2019

README.md

EGG 🐣: Emergence of lanGuage in Games

Introduction

EGG is a toolkit that allows researchers to quickly implement multi-agent games with discrete channel communication. In such games, the agents are trained to communicate with each other and jointly solve a task. Often, the way they communicate is not explicitly determined, allowing agents to come up with their own 'language' in order to solve the task. Such setup opens a plethora of possibilities in studying emergent language and the impact of the nature of task being solved, the agents' models, etc. This subject is a vibrant area of research often considered as a prerequisite for general AI. The purpose of EGG is to offer researchers an easy and fast entry point into this research area.

EGG is based on PyTorch and provides: (a) simple, yet powerful components for implementing communication between agents, (b) a diverse set of pre-implemented games, (c) an interface to analyse the emergent communication protocols.

Key features:

  • Primitives for implementing a single-symbol or variable-length communication (with vanilla RNNs, GRUs, LSTMs or Transformers);
  • Training with optimization of the communication channel with REINFORCE or Gumbel-Softmax relaxation via a common interface;
  • Simplified configuration of the general components, such as checkpointing, optimization, tensorboard support, etc;
  • Provides a simple CUDA-aware command-line tool for grid-search over parameters of the games.

To fully leverage EGG one would need at least a high-level familiarity with PyTorch. However, to get a taste of communication games without writing any code, you could try a dataset-based game, which allow you to experiment with different signaling games by simply editing input files.

Partial list of implemented games

This is a list of representative games implemented at the moment:

  • MNIST autoencoder tutorial: A Jupyter tutorial that implements a MNIST discrete auto-encoder step-by-step, covering the basic concepts of EGG. The tutorial starts with pre-training a "vision" module and builds single- and multiple symbol auto-encoder communication games with channel optimization done by Reinforce and Gumbel-Softmax relaxation (notebook / colab).
  • egg/zoo/signal_game: Modern version of a Lewis' signaling game. In this game, Sender is presented with a target image and with one or more distractor images. Then all images are shuffled and Receiver has to point to the target image based on a message from Sender. This implementation is based on Diane Bouchacourt's code.
  • egg/zoo/simple_autoenc: Discrete auto-encoder Sender/Receiver game that auto-encodes one-hot vectors using variable-length messages.
  • egg/zoo/mnist_autoenc: Discrete MNIST auto-encoder game. In this Sender/Receiver game, Sender looks onto a MNIST image and sends a single symbol to Receiver, who tries to recover the image.
  • egg/zoo/summation: Sender and Receiver are jointly trained to recognize the a^nb^n grammar: Sender reads an input sequence and Receiver answers if the sequence belongs to the grammar. Which agent actually counts, Sender or Receiver? Does Sender make the decision and send it to Receiver? Or does Sender encode the incoming sequence in the message and it is Receiver that make the decision? Or something in-between?
  • egg/zoo/external_game: A signaling game that takes inputs and ground-truth outputs from CSV files.
  • egg/zoo/objects_game: A Sender/Receiver game where the Sender sees a target as a vector of discrete properties (e.g. [2, 4, 3, 1] for a game with 4 dimensions) and the Receiver has to recognize the target among a lineup of target+distractor(s).
  • egg/zoo/language_bottleneck contains a set of games that study the information bottleneck property of the discrete communication channel. This poperty is illustrated in an EGG-based example of MNIST-based style transfer without an adversary (notebook / colab).
  • egg/zoo/channel: Similar to egg/zoo/simple_autoenc but here the inputs are drawn from a specific distribution. This game exemplifies the environment impact on the the nature of the emergent language.

We are adding games all the time: please look at the egg/zoo directory to see what is available right now. Submit an issue if there is something you want to have implemented and included.

More details on each game's command line parameters are provided in the games' directories.

An important technical point

EGG supports Reinforce and Gumbel-Softmax optimization of the communication channel. Currently, Gumbel-Softmax-based optimization is only supported if the game loss is differentiable. The MNIST autoencoder game tutorial illustrates both Reinforce and Gumbel-Softmax channel optimization when using a differentiable game loss. The signaling game has a non-differentiable game loss, and the communication channel is optimized with Reinforce.

Installing EGG

Generally, we assume that you use PyTorch 1.0.0 or newer (1.1.0 is advised) and Python 3.6 or newer.

  1. (optional) It is a good idea to develop in a new conda environment, e.g. like this:
    conda create --name egg37 python=3.7
    conda activate egg37
    
  2. EGG can be installed as a package to be used as a library
    pip install git+ssh://git@github.com/facebookresearch/EGG.git
    
    or via https
    pip install git+https://github.com/facebookresearch/EGG.git
    
    Alternatively, EGG can be cloned and installed in editable mode, so that the copy can be changed:
    git clone git@github.com:facebookresearch/EGG.git && cd EGG
    pip install --editable .
    
  3. Then, we can run a game, e.g. the MNIST auto-encoding game:
    python -m egg.zoo.mnist_autoenc.train --vocab=10 --n_epochs=50

EGG structure

The repo is organised as follows:

- tests # tests for the common components
- docs # documentation for EGG
- egg
-- core # common components: trainer, wrappers, games, utilities, ...
-- zoo  # pre-implemented games 
-- nest # a tool for hyperparameter grid search

How-to-Start and Learning more

  • The step-by-step MNIST autoencoder tutorial goes over all essential steps to create a full-featured communication game with variable length messages between the agents. NB: depending on your computational resources, this might take a while to run! (open in colab)
  • The simplest starter code for implementing a Sender/Receiver game is the MNIST autoencoder game, MNIST auto-encoder game. The game features both Gumbel-Softmax and Reinforce-based implementations.
  • EGG provides some utility boilerplate around commonly used command line parameters. Documentation about using it can be found here.
  • A brief how-to for tensorboard is here.
  • To learn more about the provided hyperparameter search tool, read this doc.
  • Our paper provides a high-level view of the toolkit and points to further resources.

Citation

If you find EGG useful in your research, please cite:

@article{Kharitonov2019,
 author = {Kharitonov, Eugene and Chaabouni, Rahma and Bouchacourt, Diane and Baroni, Marco},
 title = "{EGG: a toolkit for research on Emergence of lanGuage in Games}",
 journal={arXiv preprint arXiv:1907.00852}
}

Contributing

Please read the contribution guide.

Testing

Run pytest:

pytest

All tests should pass.

Licence

EGG is licensed under the MIT license. The text of the license can be found here.

You can’t perform that action at this time.