CR-DARTS: Channel Redistribution-based Differentiable Architecture Search

Abstract

Differentiable Architecture Search (DARTS) has shown promising results in automating the design of deep learning models. However, its search process is computationally expensive because it evaluates all candidate operations simultaneously, often leading to an over-parameterized and inefficient search network. To reduce the computational cost, DARTS employs a smaller search network than the final evaluation network, which introduces an architecture optimization gap that limits real-world performance. To overcome this limitation, we introduce CR-DARTS, a multi-stage search framework designed to bridge the architecture optimization gap through an adaptive channel redistribution strategy. CR-DARTS reduces the computational complexity of the search network by compressing the shared input features among candidate operations and restoring the network dimensions via channel-wise feature concatenation. In addition, it progressively eliminates underperforming operations and redistributes the number of channels for more relevant feature extraction, thereby narrowing the gap between the search and evaluation networks.We validated CR-DARTS on two diverse computer vision tasks to assess its generalizability. Experimental results show that the proposed search framework reduces the memory requirement of the DARTS algorithm by up to 4.3×, while addressing the architecture optimization gap. Moreover, in the evaluation phase, the discovered architecture achieves up to 25.3% reductions in computational complexity and 50.6% faster inference time compared to state-of-the-art methods, while maintaining comparable accuracy. It also produces a competitive fire segmentation network that outperforms the state-of-the-art methods while maintaining similar computational efficiency. These results demonstrate that CR-DARTS is a practical solution for neural architecture search. Source code will be made publicly available at https://github.com/Realistic3D-MIUN/CR-DARTS.