signal_game

This code implements the signalling game described in [1]. The game proceeds as follows:

• Sender is shown a target image alongside with one or many distractor images,
• Sender sends a one-symbol message to Receiver,
• Receiver obtains Sender's message and all images in random order,
• Receiver predicts which of the received images is the target one and agents are rewarded if the prediction is correct.

To run the game, you need to pre-install h5py (pip install h5py).

The game can be launched with the following command (with appropriate path to the data):

python -m egg.zoo.signal_game.train --root=/private/home/kharitonov/work/egg/data/concepts/

The data used in the paper can be downloaded from this link.

The game can be configured with the following command-line parameters:

• --root specifies the root folder of the data set
• --tau_gs set the softmax temperature for Sender; valid for both Reinforce and Gumbel-Softmax training (defaults to 10.0)
• --game_size set the number of the images used (one of them is the target, the rest are distractors, default: 2)
• --same whether the distractor images should be sampled from the same concept as the target (either 0 or 1, default: 0)
• --vocab_size sets the number of symbols for communication (default: 100)
• --batch_size sets the batch size (default: 32)
• --embedding_size sets the size of the symbol embeddings used by Receiver (default: 50)
• --hidden_size the hidden layer size used by both agents (default: 20)
• --batches_per_epoch how many batches per epoch (default: 100)
• --mode specifies which training mode will be used - either Gumbel Softmax relaxation (--mode=gs) or Reinforce (--mode=rf) (default: rf)
• --gs_tau sets the Gumbel Softmax relaxation temperature (default: 1.0)

It also accepts parameters that are common for all games, e.g.

• --n_epochs the number of training epochs to run (default: 10)
• --random_seed sets the random seed
• --lr , --optimizer, ... see this doc.

[1] "Multi-agent cooperation and the emergence of (natural) language, A. Lazaridou, A. Peysakhovich, M. Baroni [arxiv]