ABCP: Automatic Blockwise and Channelwise Network Pruning via Joint Search
ABCP: Automatic Blockwise and Channelwise Network Pruning via Joint Search
复制标题
DOI：
10.1109/tcds.2022.3230858
复制
发表时间：
2021-10
影响因子：
5
通讯作者：
Jiaqi Li;Haoran Li;Yaran Chen;Zixiang Ding;Nannan Li;Mingjun Ma;Zicheng Duan;Dong Zhao
中科院分区：
文献类型：
--
作者：
Jiaqi Li;Haoran Li;Yaran Chen;Zixiang Ding;Nannan Li;Mingjun Ma;Zicheng Duan;Dong Zhao
目前，越来越多的模型剪枝方法被提出来解决深度学习模型所需的计算机能力与资源受限设备之间的矛盾。然而，对于机器人检测等简单任务，大多数传统的基于规则的网络剪枝方法无法达到足够的压缩率和较低的精度损失，并且费时费力。在本文中，我们提出自动块式和通道式网络剪枝（ABCP），通过深度强化学习联合搜索用于机器人检测的块式和通道式剪枝动作。提出了一种联合样本算法，分别从离散和连续搜索空间同时生成每个残差块的剪枝选择和每个卷积层的通道剪枝率。最终得到兼顾模型精度和复杂度的最佳剪枝动作。与传统的基于规则的剪枝方法相比，该流程节省了人力，并实现了更高的压缩比和更低的精度损失。在移动机器人检测数据集上进行测试，剪枝后的 YOLOv3 模型节省了 99.5% 的浮点运算，减少了 99.5% 的参数，并实现了 $\boldsymbol {37.3\times }$ 的加速，而平均精度（mAP）损失仅为 2.8%。在机器人检测任务的 sim2real 检测数据集上，剪枝后的 YOLOv3 模型的 mAP 比基线模型提高了 9.6%，表现出更好的鲁棒性性能。
Currently, an increasing number of model pruning methods are proposed to resolve the contradictions between the computer powers required by the deep learning models and the resource-constrained devices. However, for simple tasks like robotic detection, most of the traditional rule-based network pruning methods cannot reach a sufficient compression ratio with low accuracy loss and are time consuming as well as laborious. In this article, we propose automatic blockwise and channelwise network pruning (ABCP) to jointly search the blockwise and channelwise pruning action for robotic detection by deep reinforcement learning. A joint sample algorithm is proposed to simultaneously generate the pruning choice of each residual block and the channel pruning ratio of each convolutional layer from the discrete and continuous search space, respectively. The best pruning action taking both the accuracy and the complexity of the model into account is obtained finally. Compared with the traditional rule-based pruning method, this pipeline saves human labor and achieves a higher compression ratio with lower accuracy loss. Tested on the mobile robot detection data set, the pruned YOLOv3 model saves 99.5% floating-point operations, reduces 99.5% parameters, and achieves $\boldsymbol {37.3\times }$ speed up with only 2.8% mean of average precision (mAP) loss. On the sim2real detection data set for robotic detection task, the pruned YOLOv3 model achieves 9.6% better mAP than the baseline model, showing better robustness performance.