This is the companion repository for our paper titled "Deep learning for time series classification: a review" published in Data Mining and Knowledge Discovery, also available on ArXiv.

Assuming you have docker installed. You can now use the docker image provided here.

Access the docker container via:

docker run --name somename --gpus all  -idt hassanfawaz/dl-4-tsc:0.3
docker exec -it somename bash

To run you will need to manually download the UCR archive into /dl-4-tsc/archives/:

cd /dl-4-tsc/archives
wget https://www.cs.ucr.edu/~eamonn/time_series_data_2018/UCRArchive_2018.zip
unzip -P $password UCRArchive_2018.zip

The password can be found here.

Now that you have the data and the code you can just run the code.

cd /dl-4-tsc
python -m main UCRArchive_2018 Coffee fcn _itr_0

You can also try and install with pip on your env.

The data used in this project comes from two sources:

• The UCR/UEA archive, which contains the 85 univariate time series datasets.
• The MTS archive, which contains the 13 multivariate time series datasets.

The code is divided as follows:

• The main.py python file contains the necessary code to run an experiement.
• The utils folder contains the necessary functions to read the datasets and visualize the plots.
• The classifiers folder contains nine python files one for each deep neural network tested in our paper.

To run a model on one dataset you should issue the following command:

python3 main.py TSC Coffee fcn _itr_8

which means we are launching the fcn model on the univariate UCR archive for the Coffee dataset (see constants.py for a list of possible options).

All python packages needed are listed in pip-requirements.txt file and can be installed simply using the pip command. The code now uses Tensorflow 2.0. The results in the paper were generated using the Tensorflow 1.14 implementation which can be found here. Using Tensorflow 2.0 should give the same results.
Now InceptionTime is included in the mix, feel free to send a pull request to add another classifier.

• numpy
• pandas
• sklearn
• scipy
• matplotlib
• tensorflow-gpu
• keras
• h5py
• keras_contrib

I added the results on the 128 datasets from the UCR archive 2018. Our results in the paper showed that a deep residual network architecture performs best for the time series classification task.

The following table contains the averaged accuracy over 10 runs of each implemented model on the UCR/UEA archive, with the standard deviation between parentheses.

The following table contains the averaged accuracy over 10 runs of each implemented model on the MTS archive, with the standard deviation between parentheses.

These results should give an insight of deep learning for TSC therefore encouraging researchers to consider the DNNs as robust classifiers for time series data.

If you would like to generate the critical difference diagrams using Wilcoxon Signed Rank test with Holm's alpha correction, check out the cd-diagram repository.

If you re-use this work, please cite:

@article{IsmailFawaz2018deep,
  Title                    = {Deep learning for time series classification: a review},
  Author                   = {Ismail Fawaz, Hassan and Forestier, Germain and Weber, Jonathan and Idoumghar, Lhassane and Muller, Pierre-Alain},
  journal                  = {Data Mining and Knowledge Discovery},
  Year                     = {2019},
  volume                   = {33},
  number                   = {4},
  pages                    = {917--963},
}

We would like to thank the providers of the UCR/UEA archive. We would also like to thank NVIDIA Corporation for the Quadro P6000 grant and the Mésocentre of Strasbourg for providing access to the cluster. We would also like to thank François Petitjean and Charlotte Pelletier for the fruitful discussions, their feedback and comments while writing this paper.