Metadata-Version: 2.1
Name: a3t
Version: 0.0.1
Summary: Implementation for A3T
Home-page: https://github.com/ForeverZyh/A3T
Author: Yuhao Zhang
Author-email: yuhaoz@cs.wisc.edu
License: MIT
Project-URL: Bug Tracker, https://github.com/ForeverZyh/A3T/issues
Platform: UNKNOWN
Classifier: Programming Language :: Python :: 3
Classifier: License :: OSI Approved :: MIT License
Classifier: Operating System :: OS Independent
Requires-Python: >=3.5
Requires-Dist: forbiddenfruit
Requires-Dist: torch (>=1.5)
Requires-Dist: tensorflow (>=2.0)
Requires-Dist: tensorflow-datasets
Requires-Dist: nltk (>=3.4)

A3T is an adversarial training technique that combines the strengths of augmentation and abstraction techniques. The key idea underlying A3T is to decompose the perturbation space into two subsets, one that can be explored using augmentation and one that can be abstracted.