/dirt-t

A DIRT-T Approach to Unsupervised Domain Adaptation (ICLR 2018)

Primary LanguagePythonMIT LicenseMIT

DIRT-T

Implementation of A DIRT-T Approach to Unsupervised Domain Adaptation (ICLR 2018).

Dependencies

numpy==1.14.1
scikit_image==0.13.1
scipy==1.0.0
tensorflow_gpu==1.6.0
tensorbayes==0.4.0

Download Data

Download scripts for MNIST and SVHN provided in ./data/.

Run code

Run VADA

python run_dirtt.py --datadir data --run 0 --src mnist --trg svhn --dirt 0

Run DIRT-T (pre-condition: run VADA first)

python run_dirtt.py --datadir data --run 0 --src mnist --trg svhn --dirt 5000

Tensorboard logs will be saved to ./log/ by default.

VADA and DIRT-T Performance

Test run of a single VADA run on MNIST -> SVHN, and using the final VADA model as initialization for 4 separate DIRT-T runs. DIRT-T has higher variance but, on expectation, improves upon VADA.

Tensorboard Visualization

VADA