Abstract: Class imbalance is a common problem that reduces the performance of classification models. One typical solution is to oversample the minority class. However, classical oversampling techniques such as SMOTE or ADASYN are ill-suited for deep learning approaches since they work in feature space. Recently, Generative Adversarial Networks (GANs) have been successfully used to generate artificial training data to re-balance datasets. Nevertheless, these approaches are data hungry and it remains a challenge to train GANs on the limited data of the minority class. In this work, we plan to leverage recent advances in data-efficient GAN training to advance the state of the art in oversampling approaches.
Dataset IR |
MNIST 10 |
50 |
100 |
Fashion-MNIST 10 |
50 |
100 |
CIFAR10 10 |
50 |
100 |
SVHN 10 |
50 |
100 |
---|---|---|---|---|---|---|---|---|---|---|---|---|
EfficentNet | - | - | - | - | - | - | - | - | - | - | - | - |
EfficentNet + Oversampling | - | - | - | - | - | - | - | - | - | - | - | - |
EfficientNet + WGAN | - | - | - | - | - | - | - | - | - | - | - | - |
EfficientNet + WGAN + ADA | - | - | - | - | - | - | - | - | - | - | - | - |
Install the required dependencies:
pip install -r requirements.txt
Run the following for training the GAN:
python main.py --config_path=configs/gan.yaml
Run the following for training the classification model:
python main.py --config_path=configs/classification.yaml