PGADA: Perturbation-Guided Adversarial Alignment for Few-shot Learning Under the Support-Query Shift

[Paper 🤗]

News

• 🔥 Release Sample Code.
• 🏆 PGADA receives Best Student Paper Award at PAKDD 2022.
• 🔥 PGADA is accepted by PAKDD 2022.

Enviroment

• GPU > 8G (>=24G for mini-Imagenet)
• Python 3.7
• Pytorch 1.7
• CUDA 10

Getting started

virtualenv venv --python=python3.7
source venv/bin/activate

Then install dependencies: pip install -r requirements.txt Some perturbations used in CIFAR-100-C-FewShot and miniImageNet-C use Wand: sudo apt-get install libmagickwand-dev

Data

To install the datasets to your machine, please follow this walkthrough.

Run an experiment

Configure your experiment by changing the values in configs/*.py, then launch your experiment. (Make Sure all config are right) python -m scripts.erm_training

Testing python -m scripts.eval_model

On some machines, the src module will not be found by Python. If this happens to you, run export PYTHONPATH=$PYTHONPATH:path/to/FewShiftBed to tell Python where you're at.

All outputs of the experiment (explicit configuration, logs, trained model state and TensorBoard logs) can then be found in the directory specified in configs/experiment_config.py. By default, an error will be risen if the specified directory already exists (in order to not harm the results of previous experiments). You may change this behaviour in configs/experiment_config.py by setting OVERWRITE = True.

Reproducing results

See the detailed documentation here.

Track trainings with Tensorboard

We log the loss and validation accuracy during the training for visualization in Tensorboard. The logs of an experiment can be found in the output directory (events.out.tfevents.[...]). To visualize them in Tensorboard, run:

tensorboard --logdir=output_dir

References

PGADA code is modified from https://github.com/ebennequin/meta-domain-shift and image perturbations are modified from https://github.com/hendrycks/robustness

Cite this Work:

@inproceedings{jiang2022pgada,
  title={PGADA: Perturbation-Guided Adversarial Alignment for Few-Shot Learning Under the Support-Query Shift},
  author={Jiang, Siyang and Ding, Wei and Chen, Hsi-Wen and Chen, Ming-Syan},
  booktitle={Advances in Knowledge Discovery and Data Mining: 26th Pacific-Asia Conference, PAKDD 2022, Chengdu, China, May 16--19, 2022, Proceedings, Part I},
  pages={3--15},
  year={2022}
}