Chen Zailong, Liu Chubo, Yang Wangdong, Li Kenli, Li Keqin
College of Information Science and Engineering, Hunan University, Hunan 410082, China.
Department of Computer Science, State University of New York, New Paltz, NY 12561, USA.
Neural Netw. 2022 Aug;152:407-418. doi: 10.1016/j.neunet.2022.05.002. Epub 2022 May 10.
Model pruning is widely used to compress and accelerate convolutional neural networks (CNNs). Conventional pruning techniques only focus on how to remove more parameters while ensuring model accuracy. This work not only covers the optimization of model accuracy, but also optimizes the model latency during pruning. When there are multiple optimization objectives, the difficulty of algorithm design increases exponentially. So latency sensitivity is proposed to effectively guide the determination of layer sparsity in this paper. We present the latency-aware automated pruning (LAP) framework which leverages the reinforcement learning to automatically determine the layer sparsity. Latency sensitivity is used as a prior knowledge and involved into the exploration loop. Rather than relying on a single reward signal such as validation accuracy or floating-point operations (FLOPs), our agent receives the feedback on the accuracy error and latency sensitivity. We also provide a novel filter selection algorithm to accurately distinguish important filters within a layer based on their dynamic changes. Compared to the state-of-the-art compression policies, our framework demonstrated superior performances for VGGNet, ResNet, and MobileNet on CIFAR-10, ImageNet, and Food-101. Our LAP allowed the inference latency of MobileNet-V1 to achieve approximately 1.64 times speedup on the Titan RTX GPU, with no loss of ImageNet Top-1 accuracy. It significantly improved the pareto optimal curve on the accuracy and latency trade-off.
模型剪枝被广泛用于压缩和加速卷积神经网络（CNN）。传统的剪枝技术只关注如何在确保模型准确性的同时去除更多参数。这项工作不仅涵盖了模型准确性的优化，还在剪枝过程中优化了模型延迟。当存在多个优化目标时，算法设计的难度会呈指数级增加。因此，本文提出了延迟敏感度，以有效地指导层稀疏度的确定。我们提出了延迟感知自动剪枝（LAP）框架，该框架利用强化学习自动确定层稀疏度。延迟敏感度被用作先验知识并融入探索循环。我们的智能体不是依赖单一的奖励信号，如验证准确性或浮点运算次数（FLOP），而是接收关于准确性误差和延迟敏感度的反馈。我们还提供了一种新颖的滤波器选择算法，以根据其动态变化准确区分层内的重要滤波器。与当前最先进的压缩策略相比，我们的框架在CIFAR-10、ImageNet和Food-101数据集上，在VGGNet、ResNet和MobileNet模型上展现出了卓越的性能。我们的LAP使得MobileNet-V1在Titan RTX GPU上的推理延迟实现了约1.64倍的加速，且不损失ImageNet Top-1准确率。它显著改善了准确性和延迟权衡上的帕累托最优曲线。
