FP-DARTS: Fast parallel differentiable neural architecture search for image classification

Neural Architecture Search (NAS) has made remarkable progress in automatic machine learning. However, it still suffers massive computing overheads limiting its wide applications. In this paper, we present an efficient search method, Fast Parallel Differential Neural Architecture Search (FP-DARTS). The proposed method is carefully designed from three levels to construct and train the super-network. Firstly, at the operation-level, to reduce the computational burden, different from the standard DARTS search space (8 operations), we decompose the operation set into two non-overlapping operator sub-sets (4 opera-tions for each). Adopting these two reduced search spaces, two over-parameterized sub-networks are constructed. Secondly, at the channel-level, the partially-connected strategy is adopted, where each sub-network only adopts partial channels. Then these two sub-networks construct a two-parallel-path super -network by addition. Thirdly, at the training-level, the binary gate is introduced to control whether a path participates in the super-network training. It may suffer an unfair issue when using softmax to select the best input for intermediate nodes across two operator sub-sets. To tackle this problem, the sigmoid func-tion is introduced, which measures the performance of operations without compression. Extensive exper-iments demonstrate the effectiveness of the proposed algorithm. Specifically, FP-DARTS achieves 2.50% test error with only 0.08 GPU-days on CIFAR10, and a state-of-the-art top-1 error rate of 23.7% on Ima-geNet using only 2.44 GPU-days for search.(c) 2022 Elsevier Ltd. All rights reserved.

Automated summary

The paper proposes an efficient search method, FP-DARTS, which is carefully designed to construct and train a super-network from three levels. Different strategies are adopted to reduce computational burden and improve efficiency. extensive experiments demonstrate the effectiveness of the proposed algorithm.