论文部分内容阅读
Pointwise convolution is usually utilized to expand or squeeze features in modern lightweight deep models. However, it takes up most of the overall computational cost(usually more than 90%). This paper proposes a novel Poker module to expand features b