This is the code repository for PAC-Bayesian Domain Adaptation Bounds for Multiclass Learners to appear in UAI 2022.
This research was conducted in conjunction with The Change that Matters in Discourse Parsing: Estimating the Impact of Domain Shift on Parser Error to appear in Findings of ACL 2022.
Some of the code is shared across repositories as detailed below.
Code for running DANN experiments and summarizing DANN results is available in this repository. Code for running and summarizing additional experiments is available in the shared code repository here. Further, a python package designed to compute key terms in the bounds we propose is available here.
Please consider citing one or both papers if you use this code.
OpenReview (UAI 2022): https://openreview.net/pdf?id=S0lx6I8j9xq
arXiv (ACL 2022): https://arxiv.org/abs/2203.11317
shared code: https://github.com/anthonysicilia/multiclass-domain-divergence
ACL code: https://github.com/anthonysicilia/change-that-matters-ACL2022
package: https://github.com/anthonysicilia/classifier-divergence
For ease of use, we have created a python script make_scripts.py to generate example bash scripts. In many cases, these may be exactly identical to the scripts used to generate results in the paper. The bash scripts interfrace with the experiments module to create the raw results for DANN experiments. Following this, code for summarizing raw results can be run using results.py. Feel free to contact us with any questions (e.g., by raising an issue here or using the contact information available in the accompanying papers).
Code was run using the following versions:
- python==3.7.4
- matplotlib==3.5.0
- numpy==1.21.2
- pandas==1.3.5
- scipy==1.7.3
- seaborn==0.12.1
- torch==1.10.2 (build py3.7_cuda10.2_cudnn7.6.5_0)
- tqdm==4.45.0
This paper is one of a series from our lab using learning theory to study understanding and generation in NLP. Check out some of our other papers here: