基于自适应阈值的循环剪枝算法

针对YOLO系列目标检测算法中复杂的网络模型和大量冗余参数问题，提出了一种基于自适应阈值的循环剪枝算法：在经过基础训练和稀疏化训练后，进入到自适应阈值剪枝模块，该模块针对缩放因子分布情况，通过缩放因子对通道和卷积层的重要性进行评估，自主学习到一个剪枝阈值，再对网络模型进行剪枝，此过程可以循环进行，并在通道剪枝和层剪枝中应用。该算法中的阈值不是人为设定，而是针对当前网络结构学习获得，通过剪枝获得一个更优的精简模型。算法实验基于YOLOv3在三个数据集上验证，结果表明，该算法对不同数据集、不同网络结构表现出较强的适应性，与传统固定阈值相比，通过自适应阈值剪枝的模型在检测精度、压缩效果、推理速度等方面都取得了更优的效果。

Adaptive threshold-based recurrent pruning algorithm

For the problem of complex network model and large number of redundant parameters in YOLO series object detection algorithm, this paper proposed an adaptive threshold recurrent pruning algorithm, after the base training and sparsity training, it entered into the adaptive threshold pruning module, which evaluated the importance of channels and convolutional layers by scaling factors for the distribution of scaling factors, and learns to a pruning threshold autonomously. Then it pruned the network model, and this process could be cyclic and applied in channel pruning and layer pruning. The thresholds in this algorithm were not artificially set, but were obtained by learning for the current network structure and obtaining a better streamlined model by pruning. It validated the algorithm experiments on three datasets based on YOLOv3, and the results show that the algorithm shows strong adaptability to different datasets and different network structures, and the model pruned by adaptive threshold which achieves better results in terms of detection accuracy, compression effect, and inference speed compared with the traditional fixed thresholds.