Domain adaptation (DA) offers a valuable means to reuse data and models for new problem domains. However, robust techniques have not yet been considered for time series data with varying amounts of data availability. In our paper, we make three main contributions to fill this gap. First, we propose a novel Convolutional deep Domain Adaptation model for Time Series data (CoDATS) that significantly improves accuracy and training time over state-of-the-art DA strategies on real-world sensor data benchmarks. By utilizing data from multiple source domains, we increase the usefulness of CoDATS to further improve accuracy over prior single-source methods, particularly on complex time series datasets that have high variability between domains. Second, we propose a novel Domain Adaptation with Weak Supervision (DA-WS) method by utilizing weak supervision in the form of target-domain label distributions, which may be easier to collect than additional data labels. Third, we perform comprehensive experiments on diverse real-world datasets to evaluate the effectiveness of our domain adaptation and weak supervision methods. Results show that CoDATS for single-source DA significantly improves over the state-of-the-art methods, and we achieve additional improvements in accuracy using data from multiple source domains and weakly supervised signals.

Preprint: https://arxiv.org/abs/2005.10996

Overview:

• Download data and convert to .tfrecord files for TensorFlow (./generate_tfrecords.sh)
• Train models (main.py or kamiak_train_*.srun)
• Evaluate models (main_eval.py or kamiak_eval_*.srun)
• Analyze results (analysis.py)

We require the following packages (module load used on Kamiak, WSU's cluster). Adjust for your computer setup.

module load cuda/10.1.105 cudnn/7.6.4.38_cuda10.1 python3/3.7.4
pip install --user --upgrade pip
export PATH="$HOME/.local/bin:$PATH"
pip3 install --user --upgrade pip
pip3 install --user --upgrade numpy cython
pip3 install --user --upgrade tensorflow-gpu pillow lxml jupyter matplotlib pandas scikit-learn scipy tensorboard rarfile tqdm pyyaml grpcio absl-py

# If using --moving_average or F1 score metrics (typically tensorflow-addons, but that errors at the moment with TF 2.2)
pip3 install --user git+https://github.com/tensorflow/addons.git@r0.9

Or, to use only the CPU, set up as follows and modify the train scripts to source kamiak_tensorflow_cpu.sh.

module load python3/3.7.4
export PATH="$HOME/.local/bin:$PATH"
pip3 install --user --upgrade virtualenvwrapper
export VIRTUALENVWRAPPER_PYTHON="$(which python3)"
mkdir -p ~/Envs
export WORKON_HOME=~/Envs
source ~/.local/bin/virtualenvwrapper.sh
mkvirtualenv -p python3 tensorflow_cpu

which pip # check it's ~/Envs/tensorflow_cpu/bin/pip
which python3 # check it's ~/Envs/tensorflow_cpu/bin/python3

# Note: we don't use --user for virtual environments
pip install --upgrade numpy cython
# Note: here it's "tensorflow" not "tensorflow-gpu" -- the rest is the same.
pip install --upgrade tensorflow pillow lxml jupyter matplotlib pandas scikit-learn scipy tensorboard rarfile tqdm pyyaml grpcio absl-py

# If using --moving_average or F1 score metrics
pip install git+https://github.com/tensorflow/addons.git@r0.9

Train a CoDATS-WS model (i.e. using DA-WS) on person 14 of the UCI HAR dataset and adapt to person 19.

python3 main.py \
    --logdir=example-logs --modeldir=example-models \
    --method=daws --dataset=ucihar --sources=14 \
    --target=19 --uid=0 --debugnum=0 --gpumem=0

Note: if passing multiple sources (comma separated, e.g. --sources=1,2,3) with DA-WS, you probably want to pass --batch_division=sources to make sure you have enough target samples in the batch to estimate the predicted label distribution (see the Appendix in the paper).

Then evaluate that model on the holdout test data, outputting the results to a YAML file.

mkdir -p results
python3 main_eval.py \
    --logdir=example-logs --modeldir=example-models \
    --jobs=1 --gpus=1 --gpumem=0 \
    --match="ucihar-0-daws-[0-9]*" --selection="best_target" \
    --output_file=results/results_example_best_target-ucihar-0-daws.yaml

Note: there are a number of other options (e.g. --ensemble=5, --moving_average), models (e.g. --model=inceptiontime), methods (e.g. --method=dann_pad), datasets (e.g. --dataset=wisdm_at), etc. implemented that you can experiment with beyond what was included in the paper.

To run all the experiments in the paper, see the Slurm scripts for training (kamiak_train_*.srun) and evaluation (kamiak_eval_*.srun). Tweak for your cluster. Then, you can run the experiments, e.g. for the single-source experiments (excluding the upper bound, which has its own scripts):

sbatch -J train kamiak_train_ssda.srun adapt
sbatch -J eval kamiak_eval_ssda.srun adapt

Then look at the resulting results/results_*.yaml files or analyze with analysis.py.

In addition, there are scripts for visualizing the datasets (datasets/view_datasets.py), viewing dataset statistics (dataset_statistics.py), and displaying or plotting the class balance of the data (class_balance.py, class_balance_plot.py).

Here is an outline of the key elements of the code.

In the paper we propose using a particular feature extractor, task classifier, and domain classifier for CoDATS. However, to support variations, the model is split into a variety of classes using inheritance.

• models.py:DannModel -- trivial class which is used in methods.py for --method=dann, just inherits from DannModelBase and CnnModelBase (which then calls FcnModelMaker to create the model since --model=fcn is the default option)
• models.py:DannModelBase -- insert the gradient reversal layer (FlipGradient) between the feature extractor and the domain classifier
• models.py:FcnModelMaker -- create the CoDATS feature extractor
• models.py:CodatsModelMakerBase -- create the CoDATS task and domain classifiers

The two most notable baseline models VRADA and R-DANN are also implemented here.

• models.py:VradaModel and models.py:RDannModel -- trivial classes used in methods.py for --method=vrada and --method=rdann respectively, just inherit from DannModelBase and RnnModelBase
• models.py:RnnModelBase -- create the VRADA or R-DANN feature extractor, task classifier, and domain classifier
• models.py:VradaFeatureExtractor -- either use a VRNN or LSTM (depending on if this is VRADA or R-DANN) followed by some dense layers
• vrnn.py:VRNN -- implementation of the variational RNN used in VRADA

In the paper we propose using a multi-source version of domain adversarial training for CoDATS and also a weak supervision method (DA-WS) for CoDATS-WS. These can be selected with --method=dann or --method=daws.

• methods.py:MethodDann - create a DannModel, handle using labeled data from multiple source domains and unlabeled data from the target domain
• methods.py:MethodDaws - inherits from MethodDann, but simulates knowing the target-domain label proportions by estimating on the source training data, also implements the DA-WS loss

Then, similarly we have the baselines --method=vrada and --method=rdann.

• methods.py:MethodRDann - same as MethodDann but uses R-DANN model instead of the CoDATS model
• methods.py:MethodVrada - inherits from MethodDann, but implements the VRNN KL divergence and negative log likelihood losses

If you use this code in your research, please consider citing our paper:

@inproceedings{wilson2020codats,
    title={Multi-Source Deep Domain Adaptation with Weak Supervision for Time-Series Sensor Data},
    author={Wilson, Garrett and Doppa, Janardhan Rao and Cook, Diane J.},
    booktitle={KDD},
    year={2020}
}