Abstract: Differentiable architecture search is a popular gradient-based method for neural architecture search, and has achieved great success in automating design of neural network architectures.