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.

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

• Training Generative Adversarial Networks with Limited Data [paper][code]
• EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks [paper][code]