Task-Aware Progressive Hardware-Friendly Pruning

Model pruning is a key technique for model lightweighting, used to compress deep learning models for deployment on resource-constrained hardware platforms. However, existing methods often struggle to effectively balance task-specific performance with hardware efficiency. In this paper, we propose a Task-aware Progressive Hardware-friendly Pruning (TPHP) method, which is a structured pruning approach that integrates task sensitivity, hardware constraints, and task characteristics. By adopting a multi-stage progressive importance evaluation mechanism that includes multiple metrics such as weight size, gradient sensitivity, activation values, and hardware efficiency, TPHP achieves an optimal balance between model accuracy and efficiency. Experiments on ShuffleNetV2 demonstrate that TPHP outperforms current model pruning methods. With a pruning rate of 75%, TPHP reduces the model size by 75.2%, accelerates inference by 69.6%, and achieves a minimal accuracy loss of only 0.8%, resulting in less accuracy degradation, better model compression, and faster inference acceleration.