/calibration-under-covariateshift

Primary LanguagePythonApache License 2.0Apache-2.0

Calibrated Predictions with Covariate Shift via Unsupervised Domain Adaptation (AISTATS20)

This code reproduces the results in the paper.

Dataset

MNIST, USPS, SVHN, and Office31 datasets are required. To this end, run initializatio scripts in data/setup. For example, to create MNIST dataset, run

python3 init_MNIST_dataset.py

Run Experiments

Experiment scripts for digit datasets and office31 dataset are located at demo/digits and demo/office31, respectively. For example, FL+IW+Temp. experiments for the shift from MNIST to USPS,

cd demo/digits/best_classifier/MNIST2USPS
ln -s ../../../../data/setup/ datasets
python3 exp_Temp_FL_IW.py

Note that the pretrained classifiers over source are included at snapshots and loaded during the experiments.