Welcome to MulDEns code base.

This is the code for our paper titled Automated Domain Discovery from Multiple Sources to Improve Zero-Shot Generalization - https://arxiv.org/abs/2112.09802

MulDEns is written on top of DomainBeD, a PyTorch suite containing benchmark datasets and algorithms for domain generalization, as introduced in (https://arxiv.org/abs/2007.01434).

Quick start

Download the datasets:

python -m domainbed.scripts.download \
       --data_dir='DATA'

Train a MULDENS ensemble with M=2 :

python3 -m domainbed.scripts.train_aug --data_dir='DATA'\
--dataset OfficeHome --test_env 0 --trial_seed 0\
--output_dir='muldens_ouputs/OfficeHome_M2/env0/trial_seed0/'\
--hparams='{"batch_size":32,"data_augmentation":1,"MULDENS_num_models":2}' 
    

Details

• In domainbed/algorithms.py MULDENS is implemented
• Through domainbed/scripts/train_aug.py and domainbed/lib/misc_aug.py we train MULDENS
• Once checkpoints are saved, if we want to load and re-evaluate, we use domainbed/scripts/eval_muldens_aug.py

Model selection criteria

We use two different model selection criteria Overall Average and Overall Ensemble. More details in paper.

Baseline Results.

We report baseline results In search of lost domain generalization paper (https://openreview.net/pdf?id=lQdXeXDoWtI) and their github repo. Full results for commit 7df6f06 in LaTeX format available here.