Emergent Communication at Scale

If running experiments, familiarize yourself with helper.py at the very least as this defines the hyperparameters for each config as provides useful options like skipping training and doing eval only.
For changing hyperparameters, please see emergent_communication_at_scale/configs/lewis_config.py for the list of hyperparameters (and their defaults).
We are not able to replicate the 50x50 experiment even with an A100 80 GiB possibly because the original experiment were run on a TPU.
If using conda/mamba, create the environment with --channel-priority flexible. This is needed due to conda-forge not having all of the necessary packages and the nvidia channel not specifying dependencies correctly.