PR-DARTS: Pruning-Based Differentiable Architecture Search

The deployment of Convolutional Neural Networks (CNNs) on edge devices is hindered by the substantial gap between performance requirements and available processing power. While recent research has made large strides in developing network pruning methods for reducing the computing overhead of CNNs, there remains considerable accuracy loss, especially at high pruning ratios. Questioning that the architectures designed for non-pruned networks might not be effective for pruned networks, we propose to search architectures for pruning methods by defining a new search space and a novel search objective. To improve the generalization of the pruned networks, we propose two novel PrunedConv and PrunedLinear operations. Specifically, these operations mitigate the problem of unstable gradients by regularizing the objective function of the pruned networks. The proposed search objective enables us to train architecture parameters regarding the pruned weight elements. Quantitative analyses demonstrate that our searched architectures outperform those used in the state-of-the-art pruning networks on CIFAR-10 and ImageNet. In terms of hardware effectiveness, PR-DARTS increases MobileNet-v2's accuracy from 73.44% to 81.35% (+7.91% improvement) and runs 3.87$\times$ faster.