miyang211 / tilt-transfer Goto Github PK

Scripts to run the TILT transfer experiments from Papadimitriou and Jurafsky 2020, Learning Music Helps You Read: Using transfer to study linguistic structure in language models https://arxiv.org/abs/2004.14601

pretrain.py trains a model on a language, assuming you have already made the corpus object wth corpora/make_corpus_object.py and have it in corpora/pickled_files. 

A simple run would be something like:
pretrain.py --data es --trial 0 --seed 400

frozen_l2_train.py trains a pretrained model on a second language, freezing everything but the embeddings

A simple run would be something like:
frozen_l2_train.py --pretrain pt --finetune es --trials 0 5

The repo uses relative paths, so that there can be a shared system of naming and saving in models/ and corpora/pickled_files to make everything easier. However, due to this, for any of this to work you must run these scripts FROM THE DIRECTORY THAT THIS README IS IN.

All other scripts should be run from the directory in which they appear, eg cd corpora/raw_to_tokens; python process_code_corpus.py rather than python corpora/raw_to_tokens/process_code_corpus.py 

The music dataset, MAESTRO, was converted from MIDI to token indices using Chris Donahue's code here https://colab.research.google.com/drive/1CCb-t-uB6ZnqUJaLzVQAylJ3Tkd3iRG4?usp=sharing.