kaniblu / glad

Global-Locally Self-Attentive Dialogue State Tracker

Faster GLAD

In terms of functionality, this fork is identical to the base code, except that the training speed has been greatly improved.

On a machine with the GTX 1080 GPU and the 6-th gen Intel CPU, the speed-up is about 7 times.

Code	Dataset	Training Time	Training Time per Epoch
Original	WoZ	58.5 minutes	70 seconds
This fork	WoZ	7.83 minutes	9.5 seconds

The speedup is achieved by the following changes:

• sequence masks are used in place of python for loops for padding -inf before applying torch.softmax. (See self-attention and attend functions in glad.py.)
• the dialogue act inference now operates on the tensor-level instead of looping through each data sample in a mini-batch. (See Model.forward method.)

Using default options, current implementation achieves 87.1% joint goal accuracy, 92.5% turn inform accuracy and 97.1% turn request accuracy.

Following is the original README file.

Global-Locally Self-Attentive Dialogue State Tracker

This repository contains an implementation of the Global-Locally Self-Attentive Dialogue State Tracker (GLAD). If you use this in your work, please cite the following

@inproceedings{ zhong2018global,
  title={ Global-Locally Self-Attentive Encoder for Dialogue State Tracking },
  author={ Zhong, Victor and Xiong, Caiming and Socher, Richard },
  booktitle={ ACL },
  year={ 2018 }
}

Install dependencies

Using Docker

docker build -t glad:0.4 .
docker run --name embeddings -d vzhong/embeddings:0.0.5  # get the embeddings
env NV_GPU=0 nvidia-docker run --name glad -d -t --net host --volumes-from embeddings glad:0.4

If you do not want to build the Docker image, then run the following (you still need to have the CoreNLP server).

pip install -r requirements.txt

Download and annotate data

This project uses Stanford CoreNLP to annotate the dataset. In particular, we use the Stanford NLP Stanza python interface. To run the server, do

docker run --name corenlp -d -p 9000:9000 vzhong/corenlp-server

The first time you preprocess the data, we will download word embeddings and character embeddings and put them into a SQLite database, which will be slow. Subsequent runs will be much faster.

docker exec glad python preprocess_data.py

The raw data will be stored in data/woz/raw of the container. The annotation results will be stored in data/woz/ann of the container.

If you do not want to build the Docker image, then run

python preprocess_data.py

Train model

You can checkout the training options via python train.py -h. By default, train.py will save checkpoints to exp/glad/default.

docker exec glad python train.py --gpu 0

You can attach to the container via docker exec glad -it bin/bash to look at what's inside or docker cp glad /opt/glad/exp exp to copy out the experiment results.

If you do not want to build the Docker image, then run

python train.py --gpu 0

Evaluation

You can evaluate the model using

docker exec glad python evaluate.py --gpu 0 --split test exp/glad/default

You can also dump a predictions file by specifying the --fout flag. In this case, the output will be a list of lists. Each ith sublist is the set of predicted slot-value pairs for the ith turn. Please see evaluate.py to see how to match up the turn predictions with the dialogues.

If you do not want to build the Docker image, then run

python evaluate.py --gpu 0 --split test exp/glad/default

Contribution

Pull requests are welcome! If you have any questions, please create an issue or contact the corresponding author at victor <at> victorzhong <dot> com.