A progressive constraint and adaptive filter pruning framework for lightweight evolutionary neural architecture search
Аннотация
Neural architecture search (NAS) has demonstrated competitive performance on various application domains, including image classification, object detection and semantic segmentation. However, many existing methods either ignore resource constraints or rely on manually designed penalties and post-hoc pruning, resulting in suboptimal performance on resource-constrained devices. In particular, static constraint handling may prematurely discard promising but over-parameterized candidates, whereas decoupled post-hoc pruning often incurs additional search cost and may undermine the performance of the searched architectures. To address these issues, we propose PCAFP-ENAS, a Progressive Constraint and Adaptive Filter Pruning framework for efficient evolutionary NAS. PCAFP-ENAS adopts a MobileNetV3-inspired lightweight search space with variable-length encoding, enabling joint optimization of network depth and width. A progressive constraint mechanism imposes time-dependent penalties on model size and FLOPs with a length-aware correction term that preserves complex candidates early and strengthens constraint pressure as evolution proceeds. Moreover, an adaptive filter pruning operator is embedded into the evolutionary loop and performs L1-norm–guided one-shot structured pruning on selected high-potential architectures, yielding compact architectures without costly iterative fine-tuning. Experiments on CIFAR-10/100 show that PCAFP-ENAS achieves competitive or superior accuracy with fewer Params and lower FLOPs than handcrafted and NAS baselines, while requiring only a few GPU days. Results on a real-world osteoporosis X-ray dataset further demonstrate competitive accuracy with smaller models, highlighting the potential of the proposed framework for resource-constrained devices.