【摘要】 为在资源受限设备中部署先进神经网络模型，提出一种基于通道和空间注意力机制的网络稀疏化剪枝训练方法，将剪枝训练过程转化为约束优化问题。将通道和空间注意力融入稀疏化剪枝训练过程，利用连续空间损失变化情况评估不同网络层重要程度，通过稀疏化训练与动态计算及更新掩码矩阵和权重矩阵完成剪枝操作。方法实验基于CIFAR10、CIFAR100数据集上进行，实验结果表明，该方法在较为复杂数据集CIFAR100上剪枝率为90%、95%、98%时，分类准确率可达到69.91%、67.15%、60.18%,与同类方法相比，在不同数据集和剪枝率的条件下仍具有较高的分类精度。更多还原

【Abstract】 To apply an efficient network model in resource constrained devices, the sparse pruning method based on the attention mechanism of spatial and channel was proposed. The pruning training process was transformed into a constrained optimization problem, and the channel and spatial attention were integrated into the sparse pruning training process, and the importance of different network layers was evaluated using the change of continuous spatial loss. The pruning operation was completed by sparsity training, dynamic calculation and updating mask matrix and weight matrix, which reduced the calculated parameters and got a simplified network model. The contrast experiment was validated with CIFAR10 and CIFAR100 datasets. When the pruning rate is 90%, 95% and 98%, the accuracy can reach 69.91%, 67.15% and 60.18%. Compared with similar methods, the method still has better accuracy under the condition of different pruning rates.更多还原