visda2019-semisupervised

We release the source code of our submission (Rank 2) for Semi-Supervised Domain Adaptation task in VisDA-2019. Details can be referred in Technical report.

Pre-trained models and extracted scores are uploaded to https://drive.google.com/open?id=1yymuaZx4h8z9f5IPhUY87oBbh5ki9Lat .

Folder structure

├── data
│   ├── clipart
│   ├── list
│   ├── painting
│   └── real
├── README.md
└── source_code
    ├── configs
    ├── datasets
    ├── evaluate_models.py
    ├── evaluate_pseudo_round_1.sh
    ├── evaluate_pseudo_round_2.sh
    ├── evaluate_pseudo_round_3.sh
    ├── evaluate_source_only.sh
    ├── evaluation.py
    ├── experiments
    ├── lib
    ├── losses
    ├── main.py
    ├── models
    ├── pretrained
    ├── __pycache__
    ├── run_pseudo_round_1.sh
    ├── run_pseudo_round_2.sh
    ├── run_pseudo_round_3.sh
    ├── run_source_only.sh
    ├── samplers
    ├── scripts
    ├── tools
    └── trainer.py

prepare data and pretrained imagenet models

1. download images to data folder and unzip to clipart, painting and real folders.
2. merge source domain and labeled target domain samples

python tools/create_balance_data.py clipart
python tools/create_balance_data.py painting

3. pretrained ImageNet models are in pretrained/checkpoints

Training and inference

1. train base models with source domain and labeled target domain samples. We will train 7 backbones including Efficient B{4,5,6,7}, se_resnext101_32x4d, senet154 and inceptionresnetv2.

bash run_source_only.sh

You can skip this step since we provide pretrained models in experiments/full_set_models.

2. inference with models in step 1

bash evaluate_source_only.sh

You can skip this step since we provide results in experiments/all_predictions/0920_noon/multi_crop_results.

3. fusion and submission

python ./tools/ensemble_models_avg.py experiments/all_predictions/0920_noon clipart ./experiments/submission/v1
python ./tools/ensemble_models_avg.py experiments/all_predictions/0920_noon painting ./experiments/submission/v1
cat ./experiments/submission/v1/clipart_pred.txt ./experiments/submission/v1/painting_pred.txt > ./experiments/submission/v1/result.txt

This submission achieves 64.3%.

4. create pseudo labels round_1

python tools/create_pseudo_labels.py 1 clipart
python tools/create_pseudo_labels.py 1 painting

You can skip this step since we provide pre-computed pseudo labels in data/list.

5. train with pseudo labels round_1

bash run_pseudo_round_1.sh

You can skip this step since we provide pretrained models in experiments/full_set_models.

6. inference with models in step 5

bash evaluate_pseudo_round_1.sh

You can skip this step since we provide results in experiments/all_predictions/0922_night/multi_crop_results.

7. fusion and submission

python ./tools/ensemble_models_avg.py experiments/all_predictions/0922_night clipart ./experiments/submission/v2
python ./tools/ensemble_models_avg.py experiments/all_predictions/0922_night painting ./experiments/submission/v2
cat ./experiments/submission/v2/clipart_pred.txt ./experiments/submission/v2/painting_pred.txt > ./experiments/submission/v2/result.txt

This submission achieves 68.8%.

8. create pseudo labels round_2

python tools/create_pseudo_labels.py 2 clipart
python tools/create_pseudo_labels.py 2 painting

You can skip this step since we provide pre-computed pseudo labels in data/list.

9. train with pseudo labels round_2

bash run_pseudo_round_2.sh

You can skip this step since we provide pretrained models in experiments/full_set_models.

10. inference with models in step 9

bash evaluate_pseudo_round_2.sh

You can skip this step since we provide results in experiments/all_predictions/0924_noon/multi_crop_results.

11. fusion and submission

python ./tools/ensemble_models_avg.py experiments/all_predictions/0924_noon clipart ./experiments/submission/v3
python ./tools/ensemble_models_avg.py experiments/all_predictions/0924_noon painting ./experiments/submission/v3
cat ./experiments/submission/v3/clipart_pred.txt ./experiments/submission/v3/painting_pred.txt > ./experiments/submission/v3/result.txt

This submission achieves 70.5%.

12. create pseudo labels round_3

python tools/create_pseudo_labels.py 3 clipart
python tools/create_pseudo_labels.py 3 painting

You can skip this step since we provide pre-computed pseudo labels in data/list.

13. train with pseudo labels round_3

bash run_pseudo_round_3.sh

You can skip this step since we provide pretrained models in experiments/full_set_models.

14. inference with models in step 13

bash evaluate_pseudo_round_3.sh

You can skip this step since we provide results in experiments/all_predictions/fusion_results. We include 7 different inference settings: 1) normal inference, 2) multi crop inference, 3) high resolution inference, 4) high resolution with no cropping, 5) normal inference with model weight average, 6) high resolution inference with model weight average, 7) high resolution with no cropping and model weight average.

15. fusion and submission

python ./tools/ensemble_models_avg.py experiments/all_predictions/fusion_results clipart ./experiments/submission/v4
python ./tools/ensemble_models_avg.py experiments/all_predictions/fusion_results painting ./experiments/submission/v4
cat ./experiments/submission/v4/clipart_pred.txt ./experiments/submission/v4/painting_pred.txt > ./experiments/submission/v4/result.txt

This submission achieves 71.35%.

16. extra fusion with prototype centers.

python tools/ensemble_models_final.py --domain clipart --model 0,1,2,3,4,5,6 --enable_prototype --prototye_acorss_model 1 --output ./experiments/submission/v5 --source_only_proto 1
python tools/ensemble_models_final.py --domain painting --model 0,1,2,3,4,5,6 --enable_prototype --prototye_acorss_model 1 --output ./experiments/submission/v5 --source_only_proto 1
cat ./experiments/submission/v5/clipart_pred.txt ./experiments/submission/v5/painting_pred.txt > ./experiments/submission/v5/result.txt

NOTE: To use this fusion, we need to extract features for each models for both source domain image and target domain images. The features are very large (around 260GB) so we do not include in this repo. If you run step14, these features will be extracted already.

This submission achieves 71.41%.

Citation

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}
}