Welcome! This repository contains code used for constructing and running the BERT-based models applied to our school reviews analysis.

Key directories and files

• src/models/base/bert_models.py - BERT models (MeanBERT and GruBERT)
• src/models/dataset.py - code for data prep
• src/models/bert_reviews.py - code for setting model hyper parameters, computing one forward pass, loss
• src/models/core/train_nn.py - wrapper class that handles training model, early stopping, etc.
• src/models/core/experiments.py - config information for running experiments (e.g. gpu allocation, parsing args, etc)
• src/sweeps/bert_reviews_sweep.py - parameter config for running sweep of experiments

Useful commands

Running IG on bert models (after initializing virtual env)

sudo bash source venv/bin/activate CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python3.6 interp/bert_interpret.py

One run (e.g. for debugging) (after initializing virtual env)

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python2 src/models/bert_reviews.py --groupname 'mn_avg_eb_meanbert' --outcome 'mn_avg_eb' CUDA_VISIBLE_DEVICES=1 PYTHONPATH=. python2 src/models/bert_reviews.py --groupname 'mn_avg_eb_robert' --outcome 'mn_avg_eb' --model_type 'robert' --hid_dim 768

Run tensorboard

tensorboard --logdir=

Print runs in sorted order according to validation loss

PYTHONPATH=. python src/models/core/experiments.py -d runs/bert_reviews/Mar11_2020/

Looking in: runs/bert_reviews/Mar11_2020/

1.3265: runs/bert_reviews/Mar11_2020/debug/hid_dim_128

Run a sweep

PYTHONPATH=. python src/models/sweeps/bert_reviews_sweep.py --outcome mn_avg_eb --groupname pred_confounds --adv_terms=perwht,perfrl

Location of runs

runs/bert_reviews/Mar08_20/testadvloss/lr0.01_hiddim64/