AdaMatch-pytorch

This is an unofficial implementation of AdaMatch: A Unified Approach to Semi-Supervised Learning and Domain Adaptation. Official code is here written by Google Research with Jax. Paper summary and video presentation are done by myself (in Korean unfortunately).

Step-by-step explanations in Colab notebooks are here.

Requirements

You can easily install all requirements by the command

pip install -r requirements.txt

Datasets

The code supports source to target domain adaptation from SVHN to MNIST (part of DigitFive dataset presented in the paper) .

Training

python main.py --uratio 3 --tau 0.9

The code includes different hyperparameters for config including

• uratio (default=3): Ratio between source and target batch size (uratio * source = target)
• tau (default=0.9): Pseudolabel threshold for Relative confidence threshold

Default all follows from the paper.

References

• AdaMatch-pytorch by zysymu

@article{berthelot2021adamatch,
  title={AdaMatch: A Unified Approach to Semi-Supervised Learning and Domain Adaptation},
  author={Berthelot, David and Roelofs, Rebecca and Sohn, Kihyuk and Carlini, Nicholas and Kurakin, Alex},
  journal={arXiv preprint arXiv:2106.04732},
  year={2021}
}