google-smart-reply-2017

Google smart reply paper (2017) implementation in tensorflow

Getting started

1. Get Ubuntu corpus dataset for testing from here

wget https://s3.amazonaws.com/ngv-public/data.zip -O data.zip

2. Unzip and move data files wherever you want.

unzip data.zip -d .

3. Install conda environment

conda create -n sr python=3.6
pip install -r requirements.txt

source activate sr

4. Update the path variables with links to the data and where you want to save model output

# main_dual_encoder_dense.py
# path params
parser.add_argument('--root_dir', default='')
parser.add_argument('--dataset_train_path', default='')
parser.add_argument('--dataset_test_path', default='')
parser.add_argument('--dataset_val_path', default='')
parser.add_argument('--vocab_path', default='')

parser.add_argument('--model_save_dir', default='')
parser.add_argument('--test_tube_dir', default='')

5. Start training

python main_dual_encoder_dense.py