subnetwork-probing

This is code for the paper:
Low-Complexity Probing via Finding Subnetworks
Steven Cao, Victor Sanh, Alexander M. Rush
NAACL-HLT 2021

Dependencies

This code was tested with python 3.6, pytorch 1.1, and pytorch-transformers 1.2.

Data

This paper uses the Universal Dependencies dataset for dependencies and part-of-speech tagging, and the CoNLL 2003 NER Shared Task for named entity recognition.

Running the code

To run the main experiment in the paper, run python main.py PATH_TO_SAVE_DIRECTORY.