Learning Spatially-Adaptive Squeeze-Excitation Networks for Few Shot Image Synthesis

Learning light-weight yet expressive deep networks for image synthesis is a challenging problem. Inspired by a more recent observation that it is the data-specificity that makes the multi-head self-attention (MHSA) in the Transformer model so powerful, this paper proposes to extend the widely adopted light-weight Squeeze-Excitation (SE) module to be spatially-adaptive to reinforce its data specificity, as a convolutional alternative of the MHSA, while retaining the efficiency of SE and the inductive bias of convolution. It proposes a spatially-adaptive squeeze-excitation (SASE) module for image synthesis task. The proposed SASE is tested in low-shot image generative learning task, and shows better performance than prior arts.