multi-domain-generalization-readme.md

Multi-domain Generalization

Prerequisites

• Create a conda virtual environment and activate it.

conda create --name uncertainty python=3.6
conda activate uncertainty

• Install Pytorch and torchvision following the official instructions, e.g.,

conda install pytorch==1.5.0 torchvision==0.6.0 cudatoolkit=10.1 -c pytorch

• Install the dependent libraries.

pip install flake8 yapf isort yacs gdown tb-nightly future scipy scikit-learn

• Setup the environment

python setup.py develop

Dataset Preparation

Download the datasets PACS (Download link) and Office-Home-DG (Download link), then place them under the directory like:

multi-domain-generalization/DATA
├── pacs
│   ├── images/
│   └── splits
└── office_home_dg
    ├── art/
    ├── clipart/    
    ├── product/    
    └── real_world/
...

Getting Started

We utilize 1 Nvidia Tesla V100 (32G) GPU for training

• For examples, you can run the following command to train models on cartoon photo sketch and test on unseen art painting.

CUDA_VISIBLE_DEVICES=0 python tools/train.py \
--root ./DATA \
--trainer Vanilla \
--uncertainty 0.5 \
--source-domains cartoon photo sketch \
--target-domains art_painting \
--seed 11 \
--dataset-config-file configs/datasets/dg/pacs.yaml \
--config-file configs/trainers/dg/vanilla/pacs.yaml \
--output-dir output/dg/pacs/uncertainty/art_painting

• You can directly run the full multi-domain generalization experiments on PACS:

sh scripts/pacs.sh

• Similarly, you can run the full multi-domain generalization experiments on Office-Home:

sh scripts/office_home.sh

The checkpoints and logs can be found at link.

Acknowledge

The implementation is based on Dassl. We thank them for their excellent projects.