Awesome

Unsupervised Domain Adaptation of Black-Box Source Models

Introduction

This is the official code of paper "Unsupervised Domain Adaptation of Black-Box Source Models" (BMVC 2021).

Results

Digits and VisDA-2017

Office31

Prepare Environment

The requirement of some main packages are:

torch==1.6.0
torchvision==0.7.0
gorilla-core==0.2.7.3
gorilla2d==0.2.8.7

Note that gorilla-core and gorilla2d are codebases developed by Gorilla-Lab-SCUT, and it is welcome to use them and give us some advices.

It is recommended to use docker image built by mine:

docker pull zhjscut/gorilla:core0.2.7.3_2d0.2.8.7

Usage

At first clone this repository:

git clone git@github.com:zhjscut/IterLNL.git

Then run a docker image (it may need some soft link to map one's dataset directory to /data):

docker run --privileged --network host --name IterLNL \
-v directory_of_cloned_code:/IterLNL \
-v /your_dataset_directory:/data \
--shm-size=2g -w /IterLNL -it cuda101-pt160-gorilla:core0.2.7.3_2d0.2.8.7 /bin/bash

After than we can start to run the experiments~

All commands of our experiments on a certain dataset are written in dataset_name.sh, such as office31.sh.

Running Experiments

At first train a model on source domain as the black box model:

CUDA_VISIBLE_DEVICES=0 python main.py configs/pretrain/office31.yaml --source A --target W --lr 0.0005

if one receive RuntimeError: Source model is trained, please run the IterLNL experiment next, it means black box model is trained.

NOTE: then one should modify source_models in solver file, for example, solver/solver_iterlnl.py, replacing the value of key Office31_A2W in source_models as the newly-trained model path.

Next running IterLNL:

CUDA_VISIBLE_DEVICES=0 python main.py configs/train/office31.yaml --source A --target W --lr 0.003 --max_epochs 30

We don't provide the checkpoints since the training of each model is quick and there are too many tasks.

Contributing

Any pull requests or issues are welcome.

Citation

If you use this benchmark in your research, please cite this project:

@article{zhang2021unsupervised,
  title={Unsupervised Domain Adaptation of Black-Box Source Models},
  author={Zhang, Haojian and Zhang, Yabin and Jia, Kui and Zhang, Lei},
  journal={arXiv preprint arXiv:2101.02839},
  year={2021}
}