Setup

• If you have already installed Torch7, please rename its folder name.

mv ~/torch ~/torch_bak

• Download Torch7

git clone https://github.com/torch/distro.git ~/torch --recursive
cd ~/torch; bash install-deps;

• Replace ~/torch/extra/cunn/lib/THCUNN/ClassNLLCriterion.cu with the one in the ./install folder.

The original ClassNLLCriterion.cu throws an error when the input is 0. We modify this file to make it accept 0.

• Install Torch7

cd ~/torch
./install.sh

• Follow the instructions (in http://torch.ch/docs/getting-started.html) to refresh your env variables.

# On Linux with bash
source ~/.bashrc

• Install dependency

luarocks install class
pip install path.py

• Pull data

python pull_data.py

Usage

• Run pretrained models

./pretrain.sh [seq2seq|seq2tree] [jobqueries|geoqueries|atis] [lstm|attention] GPU_ID

# run seq2seq without attention
./pretrain.sh seq2seq jobqueries lstm
# print results
cat seq2seq/jobqueries/dump_lstm/pretrain.t7.sample
# run seq2seq with attention
./pretrain.sh seq2seq jobqueries attention
# print results
cat seq2seq/jobqueries/dump_attention/pretrain.t7.sample

• Run experiments

./run.sh [seq2seq|seq2tree] [jobqueries|geoqueries|atis] [lstm|attention] GPU_ID

# run seq2seq without attention
./run.sh seq2seq jobqueries lstm
# print results
cat seq2seq/jobqueries/dump_lstm/model.t7.sample
# run seq2seq with attention
./run.sh seq2seq jobqueries attention
# print results
cat seq2seq/jobqueries/dump_attention/model.t7.sample

Environment

• OS: Scientific Linux 7.1
• GCC: 4.9.1 20140922 (Red Hat 4.9.1-10)
• GPU: 980 or titan x
• CUDA: 7.5
• Torch7: c0e51b98acbb54e6655343a57152b6e711ffdc2b (https://drive.google.com/file/d/0B8yp1gOBCztycW42eDNNaExWSlU/view?usp=sharing ; 1. clean.sh 2. install.sh)

The code is only tested on the above environment.