Skip to content

Latest commit

 

History

History
77 lines (62 loc) · 2.37 KB

README.md

File metadata and controls

77 lines (62 loc) · 2.37 KB
Raw

ABAS

ABAS

Code release for Adversarial Branch Architecture Search for Unsupervised Domain Adaptation.

If you use this code or the attached files for research purposes, please cite

@inproceedings{robbiano2021adversarial,
	title        = {Adversarial Branch Architecture Search for Unsupervised Domain Adaptation},
	author       = {Robbiano, Luca and Ur Rahman, Muhammad Rameez and Galasso, Fabio and Caputo, Barbara and Carlucci, Fabio Maria},
	year         = 2022,
	booktitle    = {2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
	volume       = {},
	number       = {},
	pages        = {1008--1018},
	doi          = {10.1109/WACV51458.2022.00108}
}

Software requirements

  • CUDA
  • Python 3.6 or newer
  • PyTorch 1.6 or newer
  • Other Python libraries listed in requirements.txt

Hardware requirements

  • 10 GB available on each GPU
  • Optional but strongly recommended: a cluster capable of running at least 8 parallel GPU jobs

Run experiments

To launch an ABAS run (OfficeHome, source Art, target Clipart):

./scripts/launch_slurm_stub.sh \
  --source art-oh \
  --target clipart-oh \
  --criterion 'regression(regressors/regr_no-pseudolabels_for_oh.pkl)' \
  --run-criterion 'regression(regressors/regr_for_oh.pkl)' \
  --net resnet50 \
  --da alda \
  --num-iterations 24 \
  --min-budget 2000 \
  --max-budget 6000 \
  --kill-diverging \
  --data-root /path/to/data

The script launch_slurm_stub.sh needs to be customized according to your cluster setup. A similar script can be developed for other schedulers, like PBS. Once the job is done, a result.pkl file will be produced. To analyze the results, run

./analysis.py --result experiments/your-experiment/results_file.pkl

You can test a specific configuration with

./train_model.py \
    --net resnet50 \
    --da alda \
    --gpu 0 \
    --source art-oh \
    --target clipart-oh \
    --config base.weight_da=0.88,disc.dropout=0.1,disc.hidden_size_log=10,disc.num_fc_layers=5,net.bottleneck_size_log=9 \
    --data-root /path/to/data

Contributors

License

This code and the attached files are distributed under the BSD 3-Clause license.