Gradual Domain Adaptation via Normalizing Flows

Codes for the paper "Gradual Domain Adaptation via Normalizing Flows".

Requirements

Please check the file named gdacnf_env.yml.
This file is to create the execution environment.

conda env create -n your-env-name -f cnf_env.yml

Usage

Our experiments consist of two steps as follows.

1. Fit UMAP
2. Train CNF

For starting experiments from 1st step, please download the datasets from the links listed in the Datasets section.
After downloading the datasets, run FitUMAP.py to obtain preprocessed datasets.
Since downloading takes a very long time, it is recommended to use preprocessed datasets in this supplementary material.

The usages of our experimental script are demonstrated in runExample.sh.
We conduct the experiments on our server with Intel Xeon Gold 6354 processors and NVIDIA A100 GPU , and the training of CNF take about 5 hours per dataset.

Lastly, we can use MakeFigure.ipynb to parse the experimental results and obtain the figure shown in our papers.

Datasets

Portraits
https://www.dropbox.com/s/ubjjoo0b2wz4vgz/faces_aligned_small_mirrored_co_aligned_cropped_cleaned.tar.gz?dl=0

SHIFT15M
https://github.com/st-tech/zozo-shift15m

RxRx1
We use WILDS to load pre-processed dataset.
https://wilds.stanford.edu/datasets/

Tox21
We use MoleculeNet to load pre-processed dataset.
https://moleculenet.org/