DATSA - Adversarial Domain Adaptation Network

Source Code and for the Neurocomputing ESANN 2021 Special Issue 'Domain Adversarial Tangent Subspace Alignment for Explainable Domain Adaptation'

Contact: [email protected]
If you got problems with this implementation, feel free to write me an email.

Installation

1. Download the datasets with the links below.
2. Run pip install -e . to install the DATL package.

Training

For a simple training-evaluation demo run with preset parameters, you can use the following commands for training on Office-31 A->W

Train network via cd datl && python train_datl.py --source_dir Office-31/images/amazon/ --target_dir Office-31/images/webcam
Note that source_dir and target_dir must be replaced with your dataset locations.

1. The script trains datl on amazon vs webcam given the specified image folder paths.
2. See the Args-Parser parameter description in the file for the documentation of the parameters.
3. The best model is stored in models/.

Explainability

For the explainability results on Office-31 A->W do:

1. Replace dataset paths in line 11 and 12 in study.py.
2. cd datl && python explainability.py --source amazon --target webcam
3. The Siamese Translations (STs) are stored in results
4. T-Sne plots of STs and features are stored in plots

Reproduce Performance results

For the explainability results on Office-31 do:

1. Replace dataset paths in line 11 and 12 in study.py.
2. cd datl && python study.py --dset office
   
3. Results are stored as csv file results/

Datasets

Office-31

Office-31 dataset can be found here.

Image-clef

Image-Clef dataset can be found here.