channel

channel is a Sender/Receiver game where a pair of agents is trained to encode and decode (i.e. autoencode) a one-hot vector of a fixed dimension and transmit it over channels with different properties. The vectors to be auto-encoded might come from uniform, powerlaw or some specified distribution. The communication is performed by mean of variable-length messages; the training is done by Reinforce/GS.

This code was used in the experiments of the following paper:

• Anti-efficient encoding in emergent communication. Rahma Chaabouni, Eugene Kharitonov, Emmanuel Dupoux, Marco Baroni. arxiv

The game can be run as follows:

python -m egg.zoo.channel.train --vocab_size=3 --n_features=6 --n_epoch=50 --max_len=10 --batch_size=512 --random_seed=21

The game accepts the following game-specific parameters:

• max_len -- the maximal length of the message. Receiver's output is checked either after <eos> symbol is received or after max_len symbols;
• vocab_size -- the number of unique symbols in the vocabulary (inluding <eos>!)
• sender_cell/receiver_cell -- the cells used by the agents; can be any of {rnn, gru, lstm}
• n_features -- the dimensionality of the vectors that are auto-encoded
• n_hidden -- the size of the hidden space for the RNN cells
• embed_dim -- the size of the hidden space for the RNN cells
• sender_entropy_coeff/receiver_entropy_coeff -- the regularisation coefficients for the entropy term in the loss, used to encourage exploration in Reinforce
• sender_hidden/receiver_hidden -- the size of the hidden layers for the cells
• sender_lr/receiver_lr -- the learning rates for the agents' parameters (it might be useful to have Sender's learning rate lower, as Receiver has to adjust to the changes in Sender)
• mode={gs/rf} -- training either via GS or Reinforce
• probs={p1,p2,...} or 'probs=powerlaw' select the prior distribution over concepts