compo_vs_generalization

This code implements the experiments reported in the following paper:

• Compositionality and Generalization in Emergent Languages. Rahma Chaabouni, Eugene Kharitonov, Diane Bouchacourt, Emmanuel Dupoux, Marco Baroni. ACL 2020. [arxiv]

The game can be run as follows:

python -m egg.zoo.compo_vs_generalization.train  --n_values=3 --n_attributes=5 --vocab_size=200 --max_len=2 --batch_size=5120 --sender_cell=lstm --receiver_cell=lstm --random_seed=1

Please refer to the paper to the details of the game. The hyperparameters used in the paper are provided in ./hyperparams.

The game accepts the following game-specific parameters:

• max_len -- the length of the messages.
• vocab_size -- the number of unique symbols in the vocabulary
• sender_cell/receiver_cell -- the cells used by the agents; can be any of {rnn, gru, lstm}
• n_attributes/n_values -- the number of attributes/values in the attribute/value world
• sender_emb/receiver_emb -- the size of the embeddings for Sender and Receiver
• sender_entropy_coeff -- the regularisation coefficients for the entropy term in the loss, used to encourage exploration in Reinforce (only for Sender)
• sender_hidden/receiver_hidden -- the size of the hidden layers for the cells

Reproducibility

If you want to recover results maximally close to those reported in the paper, please use EGG v1.0. This can be done by running the following command:

git checkout v1.0

In later versions of EGG, some metrics are aggregated differently, which might lead to small discrepancies.