Unsupervised Domain Generalization by Learning a Bridge Across Domains

The official code of the CVPR 2022 paper: Unsupervised Domain Generalization by Learning a Bridge Across Domains Authors: Sivan Harary, Eli Schwartz, Assaf Arbelle, Peter Staar, Shady Abu-Hussein, Elad Amrani, Roei Herzig, Amit Alfassy, Raja Giryes, Hilde Kuehne, Dina Katabi, Kate Saenko, Rogerio Feris, Leonid Karlinsky

https://arxiv.org/abs/2112.02300

Link to our Demo: https://mitibmdemos.draco.res.ibm.com/cdr/

Installation

Install Anaconda

Create Environment

conda env create -f conda_env.yml
conda activate brad

Full list of dependencies:

In case there are any problems installing the conda environment as describes above, the following is a full list of all dependecies need to run the training, testing and demo

1. pytorch (version ~ 1.8) and the corresponding torchvision
2. scikit-image
3. scikit-learn
4. tqdm
5. requests
6. jupyterlab (for demo)
7. ipywidgets (for demo)

Data Prep

Please see data_split/DATA_README.m

Downloading our pre-trained models

The model can be downloaded from: https://drive.google.com/file/d/1T7v2xwAWQGsAv11-CEwKLmUH-TmCkue9/view?usp=sharing

Run Training

To run our model first activate the conda environment:

conda activate brad

run the main training script using torch.distributed.launch:

python -m torch.distributed.launch --nproc_per_node=<NUM_GPUS> main_brad.py --data <DATA_ROOT>/clipart_train_test.txt,<DATA_ROOT>/painting_train_test.txt,<DATA_ROOT>/real_train_test.txt,<DATA_ROOT>/sketch_train_test.txt

Please see the config.py file for all available parameters or run:

python main_brad.py -h

Run Test

To run our model first activate the conda environment:

conda activate brad

Run the main test script using torch.distributed.launch:

python -m torch.distributed.launch --nproc_per_node=<NUM_GPUS> main_brad_test.py --resume <PATH_TO_TRAINED_MODEL> --src-domain <PATH_TO_SRC_DOMAIN_TXT_FILE> --dst-domain <PATH_TO_DST_DOMAIN_TXT_FILE> 

For instance, for 1-shot with source Real and target Painting use:

python -m torch.distributed.launch --nproc_per_node=<NUM_GPUS> main_brad_test.py --resume <PATH_TO_TRAINED_MODEL> --src-domain <DATA_ROOT>/real_labeled_1.txt --dst-domain <DATA_ROOT>/painting_unlabeled_1.txt 

Use the flag --classifier to choose classifier type out of [retrieval, sgd, logistic], the default is retrieval.

Run Demo

1. Make sure that the conda environment is set properly
2. Download the DomainNet Dataset
3. Download the pre-calculated features from https://drive.google.com/drive/folders/1OvowfDCNCxPCAgaOi0nVDEpiB3AF2Uut?usp=sharing
4. Run the jupyter notebook and open demo.ipynb
5. Modify the paths to the data under data_root
6. Run the demo section