MC-NAS:一种可视化模块贡献神经架构搜索方法

MC-NAS:Visual Module Contribution Neural Architecture Search Method

现有的神经架构搜索方法无法直观地将网络模型与候选模块以及模型识别准确率之间的关系展示出来;同时很多NAS方法可扩展性差,无法将其搜索策略扩展至任意搜索空间。针对上述挑战,提出了一种可视化模块贡献神经架构搜索方法。提出了模块贡献这个概念,并通过分析贡献计算过程的窘境给出了任意搜索空间下的统一采样原则,利用统一的贡献度指导原则给出了不同搜索空间的贡献度计算策略。针对特定的约束条件通过动态网络规划算法生成神经网络体系结构。大量的实验结果表明该算法在任意搜索空间中的有效性。使用CIFAR-10、CIFAR-100和ImageNet16-120数据集在NAS-Bench-201基准测试上平均准确率达到了93.33%、71.07%、42.69%。

The existing NAS methods can not directly show the relationship between network models and candidate modules and the accuracy of model recognition.At the same time,many NAS methods have poor scalability and cannot extend their search strategies to arbitrary search space.In response to the above challenges,this paper proposes a visual module contribution neural architecture search method.In this paper,the concept of module contribution is first proposed,and the unified sampling principle in arbitrary search space is given by analyzing the dilemma of the contribution calculation process.Finally,the neural network architecture is generated through a dynamic network programming algorithm for specific constraints.Extensive experimental results demonstrate the effectiveness of the proposed algorithm.Using the CIFAR-10,CIFAR-100,and ImageNet16-120 datasets,the average accuracy on the NAS-Bench-201 benchmark is 93.33%,71.07%,and 42.69%,respectively.