cs224n-win18-squad

Code for the Default Final Project (SQuAD) for CS224n, Winter 2018

Note: this code is adapted in part from the Neural Language Correction code by the Stanford Machine Learning Group.

Assignment pdf

Getting Started

Train baseline

To start training the baseline, run the following commands:

source squad/bin/activate # Remember to always activate your squad environment
cd src # Change to code directory
python main.py --experiment_name=baseline \
               --mode=train \
               --data_dir=/mnt/training/data/squad \
               --glove_path=/mnt/training/data/glove/glove.6B.100d.txt

Tracking progress in TensorBoard

tensorboard --logdir=./experiments/

and remember to ssh with tunnelling on port 6006 (from your local machine).

Inspecting Output

Once you have a trained model, you will want to see example output to help you begin to think about error analysis, and how you might improve the model. Run the following command.

python main.py --experiment_name=baseline \
               --mode=show_examples \
               --data_dir=/mnt/training/data/squad \
               --glove_path=/mnt/training/data/glove/glove.6B.100d.txt

Run official eval helper

python main.py --experiment_name=baseline \
               --mode=official_eval \
               --data_dir=/mnt/training/data/squad \
               --glove_path=/mnt/training/data/glove/glove.6B.100d.txt \
               --json_in_path=/mnt/training/data/squad/dev-v1.1.json \
               --ckpt_load_dir=../experiments/baseline/best_checkpoint/