Li Zhemin, Zuo Xiaojing, Song Yiping, Liang Dong, Xie Zheng
College of Sciences, National University of Defense Technology, 410073, Changsha, China.
Sci Rep. 2024 Dec 28;14(1):31193. doi: 10.1038/s41598-024-82562-w.
Deep Convolutional Neural Networks (DCNNs), due to their high computational and memory requirements, face significant challenges in deployment on resource-constrained devices. Network Pruning, an essential model compression technique, contributes to enabling the efficient deployment of DCNNs on such devices. Compared to traditional rule-based pruning methods, Reinforcement Learning(RL)-based automatic pruning often yields more effective pruning strategies through its ability to learn and adapt. However, the current research only set a single agent to explore the optimal pruning rate for all convolutional layers, ignoring the interactions and effects among multiple layers. To address this challenge, this paper proposes an automatic Filter Pruning method with a multi-agent reinforcement learning algorithm QMIX, named QMIX_FP. The multi-layer structure of DCNNs is modeled as a multi-agent system, which considers the varying sensitivity of each convolutional layer to the entire DCNN and the interactions among them. We employ the multi-agent reinforcement learning algorithm QMIX, where individual agent contributes to the system monotonically, to explore the optimal iterative pruning strategy for each convolutional layer. Furthermore, fine-tuning the pruned network using knowledge distillation accelerates model performance improvement. The efficiency of this method is demonstrated on two benchmark DCNNs, including VGG-16 and AlexNet, over CIFAR-10 and CIFAR-100 datasets. Extensive experiments under different scenarios show that QMIX_FP not only reduces the computational and memory requirements of the networks but also maintains their accuracy, making it a significant advancement in the field of model compression and efficient deployment of deep learning models on resource-constrained devices.
深度卷积神经网络（DCNNs）由于其高计算和内存需求，在资源受限设备上部署时面临重大挑战。网络剪枝作为一种重要的模型压缩技术，有助于在这类设备上实现DCNNs的高效部署。与传统的基于规则的剪枝方法相比，基于强化学习（RL）的自动剪枝通常通过其学习和适应能力产生更有效的剪枝策略。然而，当前的研究只设置了一个智能体来探索所有卷积层的最优剪枝率，忽略了多层之间的相互作用和影响。为了应对这一挑战，本文提出了一种采用多智能体强化学习算法QMIX的自动滤波器剪枝方法，即QMIX_FP。DCNNs的多层结构被建模为一个多智能体系统，该系统考虑了每个卷积层对整个DCNN的不同敏感度以及它们之间的相互作用。我们采用多智能体强化学习算法QMIX，其中单个智能体对系统有单调贡献，以探索每个卷积层的最优迭代剪枝策略。此外，使用知识蒸馏对剪枝后的网络进行微调可加速模型性能提升。在包括VGG - 16和AlexNet在内的两个基准DCNNs上，针对CIFAR - 10和CIFAR - 100数据集证明了该方法的有效性。在不同场景下的大量实验表明，QMIX_FP不仅降低了网络的计算和内存需求，还保持了其准确性，使其成为模型压缩领域以及在资源受限设备上高效部署深度学习模型方面的一项重大进展。
