We extend current unsupervised domain adaptation (especially domain invariant representation learning) to solve huge covariate shift betwween source and target.
python 3.9.7
implementations of domain invariant learning algo. --algo_name can switch them (so far infeasible dan_algo.py in this manner).
file name | note |
---|---|
dann_algo.py | DANNs algo https://arxiv.org/pdf/1505.07818 |
coral_algo.py | CoRAL algo https://arxiv.org/abs/1607.01719 |
dan_algo.py | DAN algo https://arxiv.org/abs/1502.02791 |
dann2D_algo.py | see our paper(TODO: Attach in the near future) |
supervised_algo.py | supervised deep learning boilerplate for comparison test |
implementations of experiment workflow (data load, preprocess, init NN, training, evaluation).
dir name | data | execution |
---|---|---|
make_moons | https://scikit-learn.org/stable/modules/generated/sklearn.datasets.make_moons.html | python -m domain-invariant-learning.experiments.make_moons.experiment |
ecodataset | https://vs.inf.ethz.ch/res/show.html?what=eco-data | git clone https://github.com/oh-yu/deep_occupancy_detection/tree/feature/JSAI run all cells of 01.ipynb - 05.ipynb python -m domain-invariant-learning.experiments.ecodataset_synthetic.experiment |
ecodataset_synthetic | see experiment.py logic | git clone https://github.com/oh-yu/deep_occupancy_detection/tree/feature/JSAI run all cells of 01.ipynb - 05.ipynb python -m domain-invariant-learning.experiments.ecodataset_synthetic.experiment |
HHAR | https://archive.ics.uci.edu/dataset/344/heterogeneity+activity+recognition | download data python -m domain-invariant-learning.experiments.HHAR.experiment |
MNIST | https://github.com/mashaan14/MNIST-M/tree/main | download data python -m domain-invariant-learning.experiments.MNIST.experiment |
implementations of networks which include layers, fit method, predict method, predict_proba method. Domain Invariant Laerning and Without Adapt and Train on Target related free params should be set here.
file name | note |
---|---|
dann.py | Figure 4: from https://arxiv.org/pdf/1505.07818 |
codats.py | Figure 3: from https://arxiv.org/pdf/2005.10996 |
danns_2d.py | see our paper(TODO: Attach in the near future) |
isih-DA.py | Algorythm 1 from https://www.jstage.jst.go.jp/article/tjsai/39/5/39_39-5_E-O41/_article/-char/ja/ |
Definition of generic functions to be called in multiple locations within the above dir structure.