Unsupervised Domain Adaptation for Time Series Classification

This repository contains the official implementation of the benchmark titled "Deep Unsupervised Domain Adaptation for Time Series Classification: a Benchmark" available on ArXiv.

This work was conducted by the team at Ericsson Research in France as part of the open source initiative.

Average rank diagrams, using accuracy on all tested algorithms and hyperparameter tuning methods. The different colors specify the employed hyperparameter tuning method: green for IWCV, blue for Source Risk and red for Target Risk.

Contributions

We welcome contributions to our benchmark.

For contributing, please follow the contribution.md guide.

Requirements

Make sure you have an env with python==3.10 and run the following:

pip install -r requirements.txt 

Run

Make sure you are inside src/, then to run a short pipeline of CoDATS model on har dataset with source=12 and target=16 follow the two commands:

Generate the config with hydra

python3 -m _utils.generate_conf 'stages/preprocess=har' 'stages/train=CoDATS' 'utils.idxExperiment=1' 'stages.tune.tuner_config.address=null' 'stages.tune.ray_config.resume=false' 'stages.tune.ray_config.num_samples=1' 'global.train_time_limit=15' 'global.tune_time_limit=30' '++stages.tune.tuner_config.hyperparam_fixed.device=cpu' 'stages.tune.ray_config.resources_per_trial.gpu=0'

Run the pipeline

bash run.sh

View result

cat output/results/dataset_name\=har/source\=12/target\=16/classifier_name\=CoDATS/metrics.json

Docker

To build the docker image:

docker build -t uda-4-tsc:latest .

To run the docker image:

docker run --name uda-4-tsc --network host --entrypoint bash -idt -v /path/to/cloned/uda-4-tsc:/tmp  uda-4-tsc:latest

Now you have the necessary env to run the code.

Alternatively you can use the already published docker image on docker hub: TODO

Results

The full results of this benchmark are available at results/.

Reference

If you re-use this work, please cite:

@misc{uda4tsc,
      title={Deep Unsupervised Domain Adaptation for Time Series Classification: a Benchmark}, 
      author={Ismail Fawaz, Hassan and Del Grosso, Ganesh and Kerdoncuff, Tanguy and Boisbunon, Aurelie and Saffar, Illyyne},
      year={2023},
      eprint={2312.09857},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

Acknowledgement

We would like to express our sincere gratitude to all individuals managing the AILAB Server within the Ericsson Re- search department. Their support and assistance have been invaluable for the completion of this research experiments.

A special acknowledgment is extended to Laetitia Chapel, Charlotte Pelletier and Romain Tavenard from the IRISA research institute, for their collaborative spirit and the resources they shared with us. Furthermore, we deeply appreciate their critical review and constructive feedback on earlier drafts of our work, which significantly enhanced the clarity and coherence of our findings.

Finally, we would also like to express our sincere gratitude to the researchers who collected, cleaned, and curated the datasets used in this benchmark. Their invaluable contribution forms the foundation of our study, and we appreciate their efforts in providing these datasets for free.

License

This work is distributed under the BSD 3-Clause License.