重新思考深度可分离卷积内核内相关如何导致改进的MobileNet We introduce blueprint separable convolutions (BSConv) as highly efficient building blocks for CNNs. They are motivated by quantitative analyses of ke