当前位置:   article > 正文

深度学习DL调参隐藏层节点数对网络性能的影响_增加隐藏节点的数量或隐藏节点的层数是否会显著提高网络的性能?

增加隐藏节点的数量或隐藏节点的层数是否会显著提高网络的性能?

这次用于实验隐藏层节点数对网络性能的影响,训练集用的是mnist的训练集的0和1,测试集用的mnist的测试集的0和1,学习率固定位0.1,batchsize=20,训练集不加噪音。得到的数据


网络结构81*1*281*2*281*5*281*10*281*20*281*60*281*100*281*200*2
训练集全样品集全样品集全样品集全样品集全样品集全样品集全样品集全样品集
测试集全测试集全测试集全测试集全测试集全测试集全测试集全测试集全测试集
学习率ret=0.1ret=0.1ret=0.1ret=0.1ret=0.1ret=0.1ret=0.1ret=0.1
batchsizez=20z=20z=20z=20z=20z=20z=20z=20
*it=10000it=10000it=10000it=10000it=10000it=10000it=10000it=10000
         
平均值0.8907990.8976820.8852180.8947020.8889310.8942050.8983680.900473
标准差0.0250110.0146670.0204370.0150170.0305960.0196690.0192980.018502
最大值0.9105960.9110690.9011350.9134340.9134340.9139070.9129610.912961
         
训练集噪音比例zx=0zx=0zx=0zx=0zx=0zx=0zx=0zx=0
测试集噪音比例zy=0zy=0zy=0zy=0zy=0zy=0zy=0zy=0


得到结论隐藏层的节点数对网络性能没有太大影响,是多少不重要,甚至很让人意外的是即便隐藏层只有1个节点网络的性能也可能很好,虽然一个节点的隐藏层很容易过拟合,但是隐藏层节点少一点无疑会极大的加快速度,综合前两次实验,很显然一个好的网络就是一个大的batchsize和一个小的隐藏层节点数同时训练集不加噪音。


具体数据


81*1*20.805109081*2*20.853832081*5*20.885525081*10*20.8789030
全样品集0.8713341全样品集0.8992431全样品集0.8647111全样品集0.8793761
全测试集0.8793762全测试集0.8978242全测试集0.8666042全测试集0.8964052
ret=0.10.9020813ret=0.10.9001893ret=0.10.8107853ret=0.10.8864713
z=200.8803224z=200.9063394z=200.9011354z=200.8921484
it=100000.8453175it=100000.8945135it=100000.8817415it=100000.90355
average0.8926216average0.9011356average0.898776average0.8666046
0.8907990.89687870.8976820.89924370.8852180.89545970.8947020.8864717
stdevp0.8935678stdevp0.9039748stdevp0.8907288stdevp0.9001898
0.0250110.91059690.0146670.89214890.0204370.89545990.0150170.8992439
max0.89971610max0.89971610max0.88363310max0.90586610
0.9105960.907285110.9110690.858562110.9011350.899716110.9134340.85998111
*0.89451312*0.90823112*0.89025512*0.90917712
zx=00.90586613zx=00.91059613zx=00.8940413zx=00.90633913
zy=00.90302714zy=00.90823114zy=00.88599814zy=00.91343414
*0.90208115*0.90018915*0.86660415*0.90302715
*0.90066216*0.90444716*0.90066216*0.9049216
*0.9096517*0.90018917*0.89687817*0.87984917
*0.90775818*0.90397418*0.8987718*0.90917718
*0.90823119*0.91106919*0.89687819*0.91296119






81*20*20.83964081*60*20.829234081*100*20.885052081*200*20.8282880
全样品集0.8779561全样品集0.8912021全样品集0.8576161全样品集0.8916751
全测试集0.8916752全测试集0.9053932全测试集0.9053932全测试集0.909652
ret=0.10.8836333ret=0.10.9006623ret=0.10.9077583ret=0.10.9091773
z=200.8949864z=200.8883634z=200.9044474z=200.9053934
it=100000.7871335it=100000.8666045it=100000.9068125it=100000.8883635
average0.9016086average0.9001896average0.9087046average0.8945136
0.8889310.86518470.8942050.89356770.8983680.90539370.9004730.9020817
stdevp0.856678stdevp0.9039748stdevp0.8883638stdevp0.9115428
0.0305960.90633990.0196690.87133490.0192980.903590.0185020.904929
max0.90917710max0.90633910max0.88836310max0.91248810
0.9134340.897824110.9139070.907285110.9129610.904447110.9129610.91296111
*0.91296112*0.91059612*0.91059612*0.88221412
zx=00.90113513zx=00.91248813zx=00.91248813zx=00.90539313
zy=00.90681214zy=00.90113514zy=00.8349114zy=00.90917714
*0.91106915*0.90633915*0.90539315*0.90870415
*0.90633916*0.87748316*0.91296116*0.90775816
*0.91248817*0.91390717*0.90681217*0.91012317
*0.90255418*0.90444718*0.90823118*0.90775818
*0.91343419*0.89356719*0.91012319*0.90728519



声明:本文内容由网友自发贡献,不代表【wpsshop博客】立场,版权归原作者所有,本站不承担相应法律责任。如您发现有侵权的内容,请联系我们。转载请注明出处:https://www.wpsshop.cn/w/盐析白兔/article/detail/314293
推荐阅读
相关标签
  

闽ICP备14008679号