tcab_benchmark

This repository contains code for running classification experiments on adversarial attacks.

Requirements

First install dependencies, eg:

conda create -n react-detection python=3.8
source activate react-detection
pip install -r requirements.txt

Then add reactdetect folder to you environment. E.g. if you are using conda

conda develop <absolute path to reactdetect>

Then, plug into react-detect/attacks_dataset folder an attacks folder in convention of the dataset generation repo, our use pre-generated attack instances here. (not required for quickstart)

Quick-start and Examples

The reactdetect folder acts as a small library that works as the building block for other modules in the repo. For a quick glance of how to train a sklearn or pytoch type model, see various examples and annotated jupyter notebook in ./reactdetect/examples.

Running Experiments at Scale

As the feature extraction process could be reused for various experiments, the repo contains another workflow for storing pre-computed features and running experiments at scale.

• First, run encode_main.py to generate sentence level feature representations at ./reprs/samplewise. (Similarly encode_main_tw.py generates token level features at ./reprs/tokenwise)

• Then run distribute_experiments.py with proper arguments, which will create train.csv and test.csv under ./detection-experiments folder.

• Then run make_experiments.py, which takes in any directory that contains train and test csvs and make them into joblib files using cached representations in ./reprs folder. You can disable distribution of token-wise features via its commandline args.

• Finally, run detection_sw.py that takes in any directory that contains pre-made train and test joblib files, which trains a LR/LGB/RF model per its input arguments; and logs models, outputs and metrics in a unique subdirectory.

Reactdetect

./reactdetect contains the building blocks for the experiments, it is formulated to work as a small library when added to PYTHONPATH.

reactdetect/
├── aggregation/        # modules that does "string to id/vector" handling, eg token/feature embedder, etc
├── allennlp_wrappers/  # relevant classes that overrides original allennlp modules, e.g. trainer that takes in python list
├── data/               # dataset readers, data loaders, etc. The classes stays as close of allennlp ones as possible
├── examples/           # examples
├── feature_extraction/ # feature extractor class and feature extaction functions
├── featu..._tokenwise/ # feature extractor class and feature extaction functions, at token level
├── utils/              # helpful file io, pandas operations and magic vars, etc.
└── models/             # base model classes, the neural base model stays as close of original allennlp one as possible