lslrh / LtC-MSDA

Implementation of Learning to Combine: Knowledge Aggregation for Multi-Source Domain Adaptation (ECCV 2020).

Learning to Combine: Knowledge Aggregation for Multi-Source Domain Adaptation

Introduction

This project is the implementation of ``Learning to Combine: Knowledge Aggregation for Multi-Source Domain Adaptation'' in PyTorch, which is accepted by ECCV 2020.

The paper is available here: arXiv

Prerequisites

• Python 3.6
• PyTorch 1.1.0 (any version higher than 0.4.0 should work)
• CUDA 9.0 & cuDNN 7.0.5

Dataset Preparation

• Digits-five
• Office-31
• DomainNet

Pre-trained Models

• Alexnet & ResNet-18,50,101

NOTE. We experimentally found that the Caffe pretrained model outperforms the PyTorch pretrained one. If you would like to evaluate our method with other backbones, a converted model from Caffe to PyTorch maybe favored.

Training

To train the baseline model without target data, simply run:

python train.py --save_model --target $target_domain$ \
                --checkpoint_dir $save_dir$

To train the full model of LtC-MSDA, simply run:

python train.py --use_target --save_model --target $target_domain$ \
                --checkpoint_dir $save_dir$

P.S. When the ``--save model'' config is active, model's parameters, global prototypes and adjacency matrix will be stored.

Test

To evaluate the LtC-MSDA model, you can run:

python test.py --target $target_domain$ --load_checkpoint $checkpoint_file$

Citation

If this work helps your research, please cite the following paper:

@inproceedings{ wang2020learning,
  title      = {Learning to Combine: Knowledge Aggregation for Multi-Source Domain Adaptation},
  author     = {Wang, Hang and Xu, Minghao and Ni, Bingbing and Zhang, Wenjun},
  booktitle  = {European Conference on Computer Vision},
  year       = {2020}
}

Also, this method has been extended into a journal work, and we will release the code of the novel ``MRF-MSDA'' in the journal version upon acceptance (most likely in a separate repository). It will be very kind of you if you can also cite our journal work:

@article{ xu2021graphical,
  title    = {Graphical Modeling for Multi-Source Domain Adaptation},
  author   = {Xu, Minghao and Wang, Hang and Ni, Bingbing},
  journal  = {arXiv preprint arXiv:2104.13057},
  year     = {2021}
}