基于强化学习和3σ准则的组合剪枝方法

徐少铭,李钰,袁晴龙

Combination pruning method based on reinforcement learning and 3σ criterion

Shao-ming XU,Yu LI,Qing-long YUAN

表 2 VGG16全连接层不同阈值下的实验结果

Tab.2 Experimental results of different thresholds on VGG16 fully connected layer

阈值 全连接层 t pa/MB af/% pp/% ∆p/% μ Fc1 0.001 9 15.32 44.28 91.88 −1.82 Fc2 0.000 2 Fc3 0.007 0 σ Fc1 0.002 5 7.04 50.75 91.97 −1.73 Fc2 0.000 9 Fc3 0.018 7 μ+σ Fc1 0.004 4 4.48 52.74 92.42 −1.28 Fc2 0.001 2 Fc3 0.024 8 μ+2σ Fc1 0.007 3 2.12 54.57 92.31 −1.39 Fc2 0.002 2 Fc3 0.043 9 μ+3σ Fc1 0.009 6 1.00 55.34 92.27 −1.43 Fc2 0.003 2 Fc3 0.063 1