Code used to run experiments for the ICLR 2023 paper "Computational Language Acquisition with Theory of Mind".
Data, as well as fine-tuned GPT-2 model used to evaluate output utterances, can be downloaded from this Zenodo link.
The script for training new speakers is ppo.py.
Example usage: python ppo.py --total-timesteps 10000000 --supervised-coef 0.01 --game-file-path data/game_file_20.pt --dev-game-file-path data/game_file_dev.pt --render-html --exp-name [experiment name] --render-every-N 1000000 --captions-file data/train_org --track --gamma 1.0 --less-logging --use-tom --beam-size 25 --sigma-decay --tom-weight 1000.0.
The script for evaluating speakers is eval.py.
Example usage: python eval.py --total-timesteps 10000 --supervised-coef 0.01 --game-file-path game_files/test/game_file_test.pt --exp-name [experiment name]--captions-file data/test_org --gamma 1.0 --less-logging --use-coco --use-tom --beam-size 25 --sigma 0.0 --seed 517 --tom-weight 1000.0 --pretrained-path [wandb path name] --track.
build_new_game_caption.py can be used to build new datasets with custom similarity scores, while listener_pretraining.py can be used to train new listeners.
The conda environment used to run the code can be built from environment-spec.txt.
Code and data were based off of Hao Zhu's InteractGym repository.