In our final project for deep learning, we experiment with Robert Schultz's paper under new datasets, GAN models, and word embeddings.
Install pytorch and python3. The following code is an example to train the model, but you should preprocess the data before training. This assumes your data is in /data if you're not using the Docker set up. Make sure to push your changes to Git before running this script!
python main.py --train-file example.train --we-file example_embeddings --model wgantwod --mode train --train-epochs 5
Alternatively, modify and run the following
./scripts/run.sh [preprocess | train | test]
Output is saved to the output folder. There is also an image option to generate a pictoral image of a given sentence.
Ensure you have gsutils installed. Put your train and dev data in GCP Storage under namespace /data. Then run ./scripts/train-cloud.sh after adjusting hyperparameters, or submit a hyperparameter job. You can view logs using the script or on the GCP platform. The final model is saved to the /models namespace.
To update the Docker image, ensure you have docker and gsutils installed. Then run ./scripts/upload-image.sh. You should not need to do this unless you need to add extra dependencies not included in requirements.txt, or if you want to avoid installing GPU dependencies if you are running the model on GPU. To avoid uploading to Google Cloud, comment out docker push.
- Add another script for hyperparameter tuning (and support in the code)