Source code accompanying our paper, "Generating Sentences by Editing Prototypes" (paper, slides).
Authors: Kelvin Guu*, Tatsunori B. Hashimoto*, Yonatan Oren, Percy Liang (* equal contribution)
NOTES:
- The instructions below are still a work in progress.
- If you encounter any problems, please open a GitHub issue, or submit a pull request if you know the fix!
- This code requires data directories that we have not uploaded yet.
- This is research code meant to serve as a reference implementation. We do not recommend heavily extending or modifying this codebase for other purposes.
If you have questions, please email Kelvin at guu.kelvin at gmail.com.
- Yelp: download from CodaLab
- One Billion Word Benchmark: download from CodaLab
Each line of each TSV file is a (prototype, revision) edit pair, separated by a tab.
-
Install Docker. If you want to use GPUs, also install nvidia-docker.
-
Download the repository and necessary data.
DATA_DIR=$HOME/neural-editor-data REPO_DIR=$HOME/neural-editor # Download repository git clone https://github.com/kelvinguu/neural-editor.git $REPO_DIR # Set up data directory mkdir -p $DATA_DIR cd $DATA_DIR # Download word vectors wget http://nlp.stanford.edu/data/glove.6B.zip # GloVe vectors unzip glove.6B.zip -d word_vectors # Download expanded set of word vectors cd word_vectors wget https://worksheets.codalab.org/rest/bundles/0xa57f59ab786a4df2b86344378c17613b/contents/blob/ -O glove.6B.300d_yelp.txt # TODO: do the same for glove.6B.300d_onebil.txt cd .. # Download datasets into data directory wget https://worksheets.codalab.org/rest/bundles/0x99d0557925b34dae851372841f206b8a/contents/blob/ -O yelp_dataset_large_split.tar.gz mkdir yelp_dataset_large_split tar xvf yelp_dataset_large_split.tar.gz -C yelp_dataset_large_split # TODO: do the same for one_billion_split # TODO: install NLTK # our code uses this variable to locate the data export TEXTMORPH_DATA=$DATA_DIR
Before you begin, be sure to set the TEXTMORPH_DATA environment variable (see "Setup" above).
Start a Docker container:
$ python run_docker.py --root --gpu $CUDA_VISIBLE_DEVICES # enter Dockerrun_docker.pypulls the latest version of our Docker image (kelvinguu/textmorph:1.2) and then starts an interactive Docker container.--root: flag to run as root inside the container (optional)--gpu $CUDA_VISIBLE_DEVICES: if you do not have a GPU, you can skip this argument.
- Inside the container,
$DATA_DIRis mounted at/dataand$REPO_DIRis mounted at/code. - The current working directory will be
/code.
Once you are inside the container, start a training run:
$ python textmorph/edit_model/main.py configs/edit_model/edit_baseline.txttextmorph/edit_model/main.pyis the main script for training an edit model.- It takes a config file as input:
configs/edit_model/edit_baseline.txt- The config file is in HOCON format.
- The script will dump checkpoints into
$DATA_DIR/edit_runs- inside
edit_runs, each training run is assigned its own folder. - The folders are numbered 0, 1, 2, ...
- inside
main.pywill complain if the Git working tree is dirty (because it logs the current commit as a record of the code's current state)- to override this, pass
--check_commit disable
- to override this, pass