Auto-DeepLab is a hierarchical neural architecture search (NAS) method specifically designed for semantic image segmentation. Unlike previous NAS approaches that focus on searching repeatable cell structures while hand-designing the outer network structure, Auto-DeepLab proposes to search both the cell and network levels simultaneously. This hierarchical approach captures the architectural variations necessary for dense image prediction, which is more challenging than image classification due to the need to maintain high spatial resolution. The method uses a continuous relaxation of discrete architectures to efficiently perform gradient-based search, achieving state-of-the-art performance on datasets such as Cityscapes, PASCAL VOC 2012, and ADE20K without any ImageNet pretraining. The best model, Auto-DeepLab, outperforms existing methods by significant margins, demonstrating the effectiveness of the proposed hierarchical architecture search framework.Auto-DeepLab is a hierarchical neural architecture search (NAS) method specifically designed for semantic image segmentation. Unlike previous NAS approaches that focus on searching repeatable cell structures while hand-designing the outer network structure, Auto-DeepLab proposes to search both the cell and network levels simultaneously. This hierarchical approach captures the architectural variations necessary for dense image prediction, which is more challenging than image classification due to the need to maintain high spatial resolution. The method uses a continuous relaxation of discrete architectures to efficiently perform gradient-based search, achieving state-of-the-art performance on datasets such as Cityscapes, PASCAL VOC 2012, and ADE20K without any ImageNet pretraining. The best model, Auto-DeepLab, outperforms existing methods by significant margins, demonstrating the effectiveness of the proposed hierarchical architecture search framework.