On the interaction between supervision and self-play in emergent communication

Code repository of the models described in the paper accepted at ICLR 2020 On the interaction between supervision and self-play in emergent communication.

Dependencies

Python

• Python>=3.6
• PyTorch>=1.2

GPU

• CUDA>=10.1
• cuDNN>=7.6

Downloading features for IBR game

The preprocssed feature files for MS COCO images/captions can be downloaded from here. The image features are obtained through a pretrained Resnet-50 model where each feature is of dimension 2048. Extract the zip file and place the folder inside the ibr_game folder.

OR game

Change the directory to or_game.

$ cd or_game

Training a population of agents using sched S2P

$ python train.py --num-compbot-samples-train 1000 --init-supervised-iters 5 --num-selfplay-iters 20 --num-supervised-iters 5 --num-iters 100 --num-encoders-train 50

where num-compbot-samples-train is the size of seed dataset $\mathcal{D}$, num-selfplay-iters is $l$, num-supervision-iters is $m$, and num-encoders-train is the number of agents in the population.

IBR game

Change the directory to ibr_game.

$ cd ibr_game

Training a single set of agents using sched S2P

$ python train.py --num_distrs 9 --num_seed_examples 10000 --s2p_schedule sched --s2p_selfplay_updates 50 --s2p_spk_updates 50 --s2p_list_updates 50 --min_list_steps 2000 --min_spk_steps 1000 --max_iters 300

where num_distrs is the total number of distractors $D$, num_seed_examples is the size of seed dataset $\mathcal{D}$, s2p_schedule is the type of S2P, s2p_selfplay_updates is $l$, and s2p_list_updates is $m$.

Finetune listener over the whole seed data

$ python finetune.py --num_seed_examples 1000 --num_total_seed_samples 10000 --num_distrs 9  ----trainpop_files <TRAINPOP_FILES>

where num_seed_examples is the number of train samples in the seed dataset $\mathcal{D}_{train}$, num_total_seed_samples is the size of the whole seed dataset $\mathcal{D}$, and <TRAINPOP_FILES> is the path to the directory where the listener parameters are stored.

Dataset & Related Code Attribution

• MS COCO is licensed under Creative Commons.
• This project is licensed under the terms of the MIT license.

Citation

If you find the resources in this repository useful, please consider citing:

@inproceedings{lowe*2020on,
    title = {On the interaction between supervision and self-play in emergent communication},
    author = {Ryan Lowe* and Abhinav Gupta* and Jakob Foerster and Douwe Kiela and Joelle Pineau},
    booktitle = {International Conference on Learning Representations},
    year = {2020},
    url = {https://openreview.net/forum?id=rJxGLlBtwH}
}