In recent years, the neural architecture search has continuously made significant progress in the field of image recognition. Among them, the differentiable method has obvious advantages compared with other search methods in terms of computational cost and accuracy to deal with image classification. However, the differentiable method is usually composed of single cell, which cannot efficiently extract the features of the network. In response to this problem, we propose a multi-level cell progressive differentiable method which allows cells to have different types according to the levels of the network. In differentiable method, the gap between the search network and the evaluation one is large, and the correlation is low. We design an algorithm to improve the distribution of architecture parameters. We also optimize the loss function and use the regularization method of additional action to improve deep network performance. The method achieves good search and classification results on CIFAR10 and ImageNet (mobile setting).

近年来，神经体系结构搜索在图像识别领域不断取得了重大进展。其中，在处理图像分类的计算成本和准确性方面，与其他搜索方法相比，可微分方法具有明显的优势。但是，可微方法通常由单个小区组成，无法有效地提取网络的特征。针对这一问题，我们提出了一种多级小区渐进可微分方法，该方法允许小区根据网络的级别具有不同的类型。在可微分方法中，搜索网络与评价之一之间的差距较大，而相关性较低。我们设计了一种算法来改善架构参数的分布。我们还优化了损失函数，并使用附加操作的正则化方法来改善深度网络性能。该方法在CIFAR10和ImageNet（移动设置）上获得了很好的搜索和分类结果。