ZLA_Pressures

Implementation of project work for AMMAC. This repository aims at replicating the analysis performed in Chaabouni et al. (2019) and applying it to Rita Chaabouni & Dupoux (2020), Luna et al. (2020) and Mordatch & Abbeel (2018).

Intended Usage

There is no command line usage implemented. Instead, training functions should be called with the intended hyperparameters. Hyperparameter selection and performance during training are logged and, together with the model weights, are serialize in the dump folder. The evaluation module provides easy access to this data, as well as conparing runs with respect to emergent word-length-distributions.

For example usage, see the test scripts on the top level of this repo.

Next Steps (old list but some are still relevant)

• fixed length setting: check if sparse rewards (reward iff ALL output values match) is better
• refine fixed length setting
  • test for alphabet size from paper
  • we need 100% (avg reward=3.0), maybe first try sparse rewards
  • "no communication" can be used for conveying information by positional encoding, shuffling message should get rid of that problem, but probably requires completely new hyperparams
• implement mixed length setting
  • RNNs for generating message as well as "reading" it
  • check if gumbel softmax is possible
  • include mixed length in evluation
• implement aux losses and check how they perform regarding task success:
  • length penalty (penalize a priori defined "no communication" token)
  • length penalty curriculum (define sucess rate (only perfect reconstructions vs. partially correct reconstructions), find nice function alpha/try and induce it from paper fig)
  • CE(y_pred; gumbel-softmax(y_pred).detach())), with y_pred being the probabilities
    • is there a hyperparam for scaling?
    • does it even work with gs?
    • include in reward baseline? (i.e. baseline(reward+loss))
  • CE(y_pred; dirichlet(old_params, argmax(message)))
    • same conciderations as above
    • look in paper to make sure that this is actually how they did it
• misc:
  • check out embedding layers (tradeoff: runtime vs. accuracy??) between vector and networks as well as between message and networks
  • Monkey typing reference distribution
  • eval plot for comparing training curves (in .json @ "log")
  • refactor main content of test.py to training/training.py
  • refactor util functions that are here and there, so they are in some common utils.py
  • add checkpoints to training in case it crashes