Neural Iterated Learning

The structure of the project is illustrated as follows:

1. Evolutions: Evolutions control the training of all generations.
2. Models: take the inputs and produce losses for updating params.
   1. Encoders: sub-module of models to encode different inputs.
   2. Decoders: sub-module of models to generate messages based on the representation from Encoders.
   3. Speakers: sub-module of models to encode inputs and generate messages.
   4. Listeners: sub-module of models to finish the game.
   5. Losses: sub-module of models to gain loss for training models.
   6. Pipelines: a pipeline model of all above modules.
3. DataIterators: provide data to models under evolution.
   1. Prepocesses: sub-module of DataIterator to provide preprocessing functions.
   2. Voc: sub-module of DataIterator to provide dictionaries.
4. Utils: Other functions to support evolutions.