KunalSinha1125 / amortized-rsa

Learning to refer informatively by amortizing pragmatic reasoning

amortized-rsa

Learning to refer informatively by amortizing pragmatic reasoning

• shapeworld.py generates the shapeworld dataset
• models.py contains code for the models listed in the paper
  • /models/ contains pretrained models
• train.py file can be used to train the models with the following arguments:
  • --dataset specifies the dataset to train on (shapeworld or colors)
    • --generalization specifies the generalization type if training the model to generalize to new colors (new_color), combinations (new_combo), or shapes (new_shape)
  • --vocab flag generates a new vocab file
  • --s0 flag trains a literal speaker
  • --l0 flag trains a literal listener
  • --amortized flag trains an amortized speaker
    • --penalty specifies the utterance cost function (length)
      • --lmbd specifies the cost function parameter