This code implements the experiments reported in the following paper:

• Defending Compositionality in Emergent Languages. Michal Auersperger, Pavel Pecina. NAACL SRW 2022. [arxiv]

Edit

The code has been merged to the official EGG toolkit repository. You can find it along other experiments in the egg/zoo directory: egg/zoo/compo_vs_generalization_ood and ignore this repository.

---

Installation

The code uses the EGG toolkit, which can be installed like this:

pip install git+ssh://[email protected]/facebookresearch/EGG.git

Running experiments

A single experiment of the full communication game can be run as follows:

python train  --n_values=50 --n_attributes=2 --vocab_size=50 --max_len=3  --receiver=ModifReceiver --sender=ModifSender --hidden=50 --batch_size=64 --random_seed=1

To run an experiment with a single agent on half of the problem only (i.e., a learning alone experiment from the paper), run e.g.:

python -m learning_alone.train  --n_values=50 --n_attributes=2 --vocab_size=50 --max_len=5 --archpart=sender --model=OrigSenderDeterministic --hidden=50 --batch_size=64 --random_seed=1 

See the article for further details.

The scripts log information to STDOUT.

Hyperparameters

To replicate the full experiments, use the hyperparameters in ./hyperparams/modified_arch.json and ./hyperparams/orig_arch.json.

To replicate the learning alone experiments, use the hyperparameters in ./hyperparams/learning_alone/receiver.json and ./hyperparams/learning_alone/sender.json.

We use a simple gs.py script to process the hyperparameter jsons. This selects a specific configuration of the hyperparameters from a hyperparams file based on a run_id that we get from the job scheduler at our computational cluster. E.g.,

python gs.py --gs_file learning_alone/sender.json --count_runs

returns the number of runs defined by the json (40) and

python gs.py --gs_file learning_alone/sender.json --exp_type learning_alone --run_id 1

runs the first one.

Processing the log files

For convenience, we attach the log files of our full experiment runs in results/orig_arch/220520T115546 and results/modified_arch/220517T231916. The notebook ntb-results-full.ipynb processes the logs and produces Table 2 and Figure 1 from the paper.