https://github.com/tomekkorbak/compositional-communication-via-template-transfer

Code acompanying the paper Developmentally motivated emergence of compositional communication via template transfer
https://github.com/tomekkorbak/compositional-communication-via-template-transfer

communication compositionality computational-linguistics deep-learning emergent-communication gumbel-softmax language multi-agent signaling-games transfer-learning

Last synced: 7 days ago
JSON representation

Code acompanying the paper Developmentally motivated emergence of compositional communication via template transfer

• Host: GitHub
• URL: https://github.com/tomekkorbak/compositional-communication-via-template-transfer
• Owner: tomekkorbak
• Created: 2019-10-02T23:12:26.000Z (about 5 years ago)
• Default Branch: master
• Last Pushed: 2022-12-08T06:40:39.000Z (about 2 years ago)
• Last Synced: 2024-10-27T16:57:01.191Z (about 2 months ago)
• Topics: communication, compositionality, computational-linguistics, deep-learning, emergent-communication, gumbel-softmax, language, multi-agent, signaling-games, transfer-learning
• Language: Python
• Size: 23.3 MB
• Stars: 0
• Watchers: 2
• Forks: 2
• Open Issues: 12
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
Compositional communication via template transfer

==================================

This repo contains the code accompanying the paper *Developmentally motivated emergence of compositional communication via template transfer* accepted for NeurIPS 2019 workshop [Emergent Communication: Towards Natural Language](https://sites.google.com/view/emecom2019/home?authuser=0).

## Running the code

We assume Python <= 3.6. To reproduce the results of template transfer run:

```bash

pip install -r requirements.txt

unzip data.zip

python -m template_transfer.train

```

To reproduce reported baselines, run the following commands:

* Random baseline: `python -m template_transfer.train --sender_lr 0 --receiver_lr 0 --no_transfer`

* Same architecture without template transfer: `python -m template_transfer.train --no_transfer`

* Obverter: `python -m obverter.train`

Use `--help` flag for available arguments. All arguments default to hyperparameters used in the paper. I use [Neptune.ml](https://neptune.ml/) for experiment management, which is turned off by default. Pass `--neptune_project ` and set environmental variable `NEPTUNE_API_TOKEN` to log metrics using Neptune.

In case of questions, create an issue or contact Tomek Korbak under [email protected].

## Citing

```latex

@article{korbak_template_transfer_2019,

  author    = {Korbak, Tomasz and

               Zubek, Julian and

               Kuci\'{n}ski, \L{}ukasz and

               Mi\l{}oś, Piotr and

               R\k{a}czaszek-Leonardi, Joanna},

  title     = {Developmentally motivated emergence of compositional communication via template transfer},

  journal   = {NeurIPS 2019 workshop Emergent Communication: Towards Natural Language},

  year      = {2019},

  url       = {}

}

```