pholur / situation-modeling

The NP2IO Codebase

NP2IO - Threat Detection in Noun Phrases

Insider-Outsider Classification - A new paradigm for Identifying Conspiracy Theories

Pointers to get started:

1. imports.py: Change the variables here to access the model checkpoints, dataset, and intermediate pkl files available on OSF. Pay particular attention to *DATA_PATH as well as the OPT option to switch between training and inference.
2. Fine-tuning NP2IO is performed on 1-2 Titan RTX GPUs for about 2-3 hours with one of the GPUs solely attributed to data augmentation via BERT. Set REEXTRACT and SAVE to False to use a single-GPU for training post-data augmentation. The post-data augmented input files are available on OSF.
3. To train, run python train.py. TODO: Dockerize the environment. Till then transformers, pytorch1.7 is sufficient with additional minor dependencies.
4. To test, run python SCRIPT_test_against_baselines.py. This call takes a while - currently does not generate real-time outputs so it is a little annoying to run.
5. test_inline.ipynb provides a table-top setup for qualitative eval.
6. Other files are helpers.

Contact me if you have any questions.