We optimize depthwise-separable convolution through a discrete ODE module for building lightweight models, achieving remarkable performances with less parameters.