HyperTransformer: Attention-Based CNN Model Generation from Few Samples

In this work we propose a HyperTransformer, a transformer based model that generates all weights of a CNN model directly from the support samples. This approach allows to use a high-capacity model for encoding task-dependent variations in the weights of a smaller model. We show for multiple few-shot benchmarks with different architectures and datasets that our method beats or matches that of the traditional learning methods in a few-shot regime. Specifically, we show that for very small target models, our method can generate significantly better performing models than traditional few-shot learning methods. For larger models we discover that applying generation to the last layer only, allows to produce competitive or better results while being end-to-end differentiable. Finally, we extend our approach to semi-supervised regime utilizing unlabeled samples in the support set and further improving few-shot performance in the presence of unlabeled data.