A repository for code producing one of our augmentation analysis results. The code should be adaptable to all
other augmentations supported by nlpaug, though only word substitution with BERT is implemented. Our final result is
the following plot:
A given training run can be completed using python train.py runs/[some run config]. runner.py is a utility
provided to run these in a queue given a fixed number of GPUs.
eval_affinity_diversity.py is used to log the one result not included at training time, a clean-trained model's test
accuracy on augmented data.
export_csv.py downloads and processes the results from Weights & biases.
Our wandb runs can be viewed here
generate_plot_from_runs.py constructs the above plot from this data.