翻译：MAE-NAS使用定制的进化算法检测主干的多尺度微分熵。为了提高进化效率提出了一种由粗到细的策略逐步减少搜索空间。首先随机生成N个网络架构来填充种群p。一个网络架构Ft由一系列构建块组成例如ResNet块。然后我们随机选择一个块用它的变异版本替换它。在早期阶段使用粗突变在T 2 迭代之后切换到精细突变。在粗突变中块类型、内核大小、深度和宽度是随机突变的。在精细突变中只有内核大小和宽度发生突变

MAE-NAS uses a customized evolutionary algorithm to detect the multi-scale differential entropy of the backbone. To improve the efficiency of evolution, a strategy from coarse to fine is proposed to gradually reduce the search space. Firstly, N network architectures are randomly generated to fill the population p. A network architecture Ft is composed of a series of building blocks, such as ResNet blocks. Then, we randomly select a block and replace it with its mutated version. Coarse mutation is used in the early stage, and fine mutation is switched after T/2 iterations. In coarse mutation, block type, kernel size, depth, and width are randomly mutated. In fine mutation, only kernel size and width undergo mutation