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

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

• Digits-five
• Office-31
• DomainNet

• 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.

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.

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

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

If this work helps your research, please cite the following paper (it will be updated when the ECCV paper is publicized).

@article{wang2020learning,
  author    = {Hang Wang, Minghao Xu, Bingbing Ni, Wenjun Zhang},
  title     = {Learning to Combine: Knowledge Aggregation for Multi-Source Domain Adaptation},
  journal   = {CoRR},
  volume    = {abs/2007.08801},
  year      = {2020}
}