We release the source code of our submission (Rank 1) for Multi-Source Domain Adaptation task in VisDA-2019. Details can be referred in Technical report.

All the pretrained models, synthetic data generated via CycleGAN , and submission files can be downloaded from the link.

You may need a machine with 4 GPUs and PyTorch v1.1.0 for Python 3.

1. Go to the Adapt folder

2. Train source only models

bash experiments/<DOMAIN>/<NET>/train.sh

Where <DOMAIN> is clipart or painting, <NET> is the network (e.g. senet154)

Then repeat the following procedures 4 times.

bash experiments/<DOMAIN>/<NET>_<phase_id>/train.sh

1. Copy the adaptation models to the folder ExtractFeat/experiments/<phase_id>/<DOMAIN>/<NET>/snapshot

2. Extract features by running the scripts

bash experiments/<phase_id>/<DOMAIN>/scripts/<NET>.sh

3. Copy the features from experiments/<phase_id>/<DOMAIN>/<NET>/<NET>_<source_and_target_domains>/result to dataset/visda2019/pkl_test/<phase_id>/<DOMAIN>/<NET>

1. Go to the FeatFusionTest folder

2. Train feature fusion based adaptation module

bash experiments/<phase_id>/<DOMAIN>/train.sh

3. Copy the pseudo labels file to Adapt/experiments/<DOMAIN>/<NET>_<next_phase_id> for the next adaptation.

Please cite our technical report in your publications if it helps your research:

@article{pan2019visda,
  title={Multi-Source Domain Adaptation and Semi-Supervised Domain Adaptation with Focus on Visual Domain Adaptation Challenge 2019},
  author={Pan, Yingwei and Li, Yehao and Cai, Qi and Chen, Yang and Yao, Ting},
  booktitle={Visual Domain Adaptation Challenge},
  year={2019}
}

Thanks to the domain adaptation community and the contributers of the pytorch ecosystem.

Pytorch pretrained-models Cadene and EfficientNet