Multi-agent guessing game playground. Based on Multi-Agent Cooperation and the Emergence of (Natural) Language (Lazaridou et al., 2017). The experimental part of my master thesis.
Train the models using the following command:
python train.py [settings-train.yml]Experiment settings and model hyperparameters can be specified via the settings file only.
It is possible to queue up several experiments using a csv file:
python train.py [settings-train.yml] [batch-settings.csv]Final training stats of each model are written to the specified results file.
To generate a test set, run:
python make_test.py [settings-make-test.yml]To test your models, run the following command:
python test.py [settings-test.yml]The models and the test file need to be specified in the settings file.
Result analyses can be carried out via the jupyter notebooks in the analysis folder.
- plot the 6switch models
- pick the best one from each setting, move to sep folder
- test the best on the regular testset
- test the best on the same-synset testset
- cluster analysis of test output
- symbol purity of test output
- cluster analysis of embedding layer
- qualitative analysis of clusters