神经网络的分类性能是否随着卷积核数量的增加而单调的递增?
本文制作了一个3分类的网络,设置卷积核的数量分别等于1-24个,观察网络的平均分类准确率随着卷积核数量的增加到底是如何变化的。
网络结构
Mnist(0,1,2)---con3*3*n---7*7*n---30---3----(1,0,0) || (0,1,0) || (0,0,1)
将mnist的0,1,2数据集的图片处理成9*9(每隔3个点取一个点),用n个3*3的卷积核,三层网络的结构是(49*n)*30*3个节点。
用网络输出值与目标值的差值作为网络的收敛结束标准:
|输出值-目标值|=δ
让δ分别等于1e-5到1e-4.每个δ收敛199次取平均值,共收敛了24*10*199次。分别记录对应每个δ的平均准确率,最大准确率,迭代次数和收敛时间。
平均准确率
无核 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | 24 | |
平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | 平均准确率p-ave | |
1.00E-04 | 0.9827705 | 0.9753358 | 0.9804983 | 0.9811083 | 0.9811163 | 0.9818204 | 0.9825422 | 0.9831043 | 0.9832368 | 0.9831634 | 0.9829973 | 0.9837925 | 0.9831697 | 0.9835051 | 0.98385 | 0.9841039 | 0.9833406 | 0.9836009 | 0.9836951 | 0.9834156 | 0.9831985 | 0.983529 | 0.9834252 | 0.9830915 | 0.9825917 |
9.00E-05 | 0.9831266 | 0.9761263 | 0.9802412 | 0.9813733 | 0.9826029 | 0.9818604 | 0.9831458 | 0.9832304 | 0.9837222 | 0.9841645 | 0.9838516 | 0.9842827 | 0.98343 | 0.9845158 | 0.9838771 | 0.9839538 | 0.9840448 | 0.9841518 | 0.9843865 | 0.9832576 | 0.9835657 | 0.9833454 | 0.9829606 | 0.9831618 | 0.9837398 |
8.00E-05 | 0.9833837 | 0.9765526 | 0.9802029 | 0.9811099 | 0.9821909 | 0.98263 | 0.9833614 | 0.9834588 | 0.9837462 | 0.9841342 | 0.9839697 | 0.9846835 | 0.984313 | 0.9843562 | 0.984321 | 0.9842332 | 0.9843562 | 0.9842444 | 0.9847122 | 0.9834811 | 0.9841789 | 0.9835226 | 0.984131 | 0.9839793 | 0.9835083 |
7.00E-05 | 0.9837717 | 0.9765654 | 0.9811322 | 0.9815586 | 0.9826572 | 0.9834444 | 0.9840991 | 0.9838292 | 0.9843354 | 0.9845111 | 0.984353 | 0.9840512 | 0.9844584 | 0.9844376 | 0.9843498 | 0.9849134 | 0.9845047 | 0.984361 | 0.9841071 | 0.9842795 | 0.9839809 | 0.9843514 | 0.9844855 | 0.9841198 | 0.9837302 |
6.00E-05 | 0.9843242 | 0.9775235 | 0.9814165 | 0.9831186 | 0.9833326 | 0.9840432 | 0.984361 | 0.9841645 | 0.9844999 | 0.9847617 | 0.9844328 | 0.9851625 | 0.984939 | 0.9848783 | 0.9849214 | 0.9844328 | 0.9851226 | 0.9848687 | 0.9846532 | 0.9846532 | 0.9844025 | 0.9848751 | 0.9844296 | 0.9845669 | 0.9842172 |
5.00E-05 | 0.9848608 | 0.9778316 | 0.9818875 | 0.9827035 | 0.9838452 | 0.9846627 | 0.9847458 | 0.9852663 | 0.9852073 | 0.9847985 | 0.9853462 | 0.9856464 | 0.985739 | 0.9854532 | 0.9856496 | 0.9844248 | 0.9844472 | 0.9854436 | 0.98545 | 0.9853829 | 0.9850396 | 0.9844823 | 0.9846212 | 0.9843562 | 0.9844663 |
4.00E-05 | 0.985426 | 0.9786636 | 0.9824256 | 0.9841071 | 0.9846915 | 0.9848719 | 0.985715 | 0.9860136 | 0.985723 | 0.9858108 | 0.9857358 | 0.9863953 | 0.985755 | 0.9863777 | 0.9858045 | 0.9852647 | 0.9854755 | 0.9854867 | 0.9855985 | 0.985624 | 0.9857406 | 0.985228 | 0.98545 | 0.9853302 | 0.985739 |
3.00E-05 | 0.9859418 | 0.9786843 | 0.9835178 | 0.9850955 | 0.9860791 | 0.9863937 | 0.9868184 | 0.9868312 | 0.9871442 | 0.9865055 | 0.986852 | 0.9864639 | 0.986836 | 0.9868152 | 0.9867721 | 0.9866795 | 0.9866635 | 0.9866188 | 0.9866252 | 0.9865118 | 0.9865166 | 0.9858891 | 0.9859306 | 0.9862835 | 0.9860472 |
2.00E-05 | 0.9869063 | 0.9805414 | 0.9840049 | 0.9861653 | 0.987141 | 0.9872751 | 0.9879074 | 0.988107 | 0.9882875 | 0.9882108 | 0.9880176 | 0.9877557 | 0.9874635 | 0.9876072 | 0.9879362 | 0.9876376 | 0.9877318 | 0.9872512 | 0.9875881 | 0.9872496 | 0.9872559 | 0.9871394 | 0.9872895 | 0.987034 | 0.98705 |
1.00E-05 | 0.9881518 | 0.9821957 | 0.9861542 | 0.9877238 | 0.9883146 | 0.989006 | 0.989006 | 0.9889837 | 0.9893669 | 0.9891753 | 0.9894531 | 0.9893717 | 0.9893494 | 0.989204 | 0.9894867 | 0.9894867 | 0.9893589 | 0.9893957 | 0.9889278 | 0.9889645 | 0.9888815 | 0.9887282 | 0.9884232 | 0.9884615 | 0.9886324 |
从图中可以发现没有卷积核的网络也就是81*30*3的网络的分类性能大致和6核网络的性能相当,也就是如果卷积核数量少于6个不如不使用卷积核。
但是网络性能并没有随着卷积核的数量的增加而单调的增加,
这点从上图很容易观察到,平均性能的峰值顶点在第8个核附近0.989366917216081,卷积核的数量超过8个核以后性能随着卷积核的数量的增加而下降。
也就是希望通过增加卷积核来无限制提升网络性能的办法是不可行的,因为这是一个先上升后下降的趋势。
1<2<3<4<5<6=无核<7<8>9>10>11>12>13>14>15>16>17>18>19>20>21>22>23>24
再来观察最大性能
无核 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | 24 | |
最大值p-max | 最大值p-max | 最大值p-max | 最大值p-max | 最大值p-max | 最大值p-max | 最大值p-max | 最大值p-max | 最大值p-max | 最大值p-max | 最大值p-max | 最大值p-max | 最大值p-max | 最大值p-max | 最大值p-max | 最大值p-max | 最大值p-max | 最大值p-max | 最大值p-max | 最大值p-max | 最大值p-max | 最大值p-max | 最大值p-max | 最大值p-max | 最大值p-max | |
1.00E-04 | 0.9863362 | 0.9831586 | 0.9882428 | 0.9901493 | 0.9904671 | 0.9901493 | 0.9907849 | 0.9901493 | 0.9911026 | 0.9898316 | 0.9904671 | 0.9907849 | 0.9904671 | 0.9901493 | 0.9904671 | 0.9898316 | 0.9901493 | 0.9904671 | 0.9901493 | 0.9901493 | 0.9901493 | 0.9907849 | 0.9901493 | 0.9904671 | 0.9895138 |
9.00E-05 | 0.9863362 | 0.9828408 | 0.9904671 | 0.9895138 | 0.9907849 | 0.9907849 | 0.9907849 | 0.9914204 | 0.9911026 | 0.9907849 | 0.9904671 | 0.9904671 | 0.9904671 | 0.9911026 | 0.9904671 | 0.9907849 | 0.9904671 | 0.9901493 | 0.9904671 | 0.9904671 | 0.9904671 | 0.9901493 | 0.9901493 | 0.9895138 | 0.9911026 |
8.00E-05 | 0.9863362 | 0.9837941 | 0.9891961 | 0.9904671 | 0.9907849 | 0.9911026 | 0.9907849 | 0.9907849 | 0.9904671 | 0.9911026 | 0.9907849 | 0.9914204 | 0.9904671 | 0.9904671 | 0.9904671 | 0.9904671 | 0.9904671 | 0.9901493 | 0.9904671 | 0.9911026 | 0.9911026 | 0.9901493 | 0.9904671 | 0.9901493 | 0.9901493 |
7.00E-05 | 0.986654 | 0.9847474 | 0.9888783 | 0.9914204 | 0.9907849 | 0.9914204 | 0.9907849 | 0.9917382 | 0.9926915 | 0.9907849 | 0.9914204 | 0.9920559 | 0.9911026 | 0.9911026 | 0.9911026 | 0.9911026 | 0.9914204 | 0.9907849 | 0.9907849 | 0.9904671 | 0.9907849 | 0.9901493 | 0.9901493 | 0.9911026 | 0.9904671 |
6.00E-05 | 0.9869717 | 0.9844296 | 0.9907849 | 0.9911026 | 0.9907849 | 0.9907849 | 0.9914204 | 0.9920559 | 0.9930092 | 0.9914204 | 0.9917382 | 0.9904671 | 0.9914204 | 0.9907849 | 0.9911026 | 0.9914204 | 0.9914204 | 0.9907849 | 0.9904671 | 0.9914204 | 0.9914204 | 0.9901493 | 0.9911026 | 0.9904671 | 0.9907849 |
5.00E-05 | 0.9872895 | 0.9863362 | 0.9895138 | 0.9911026 | 0.9911026 | 0.9917382 | 0.9917382 | 0.9917382 | 0.9920559 | 0.9914204 | 0.9914204 | 0.9914204 | 0.9917382 | 0.9917382 | 0.9914204 | 0.9907849 | 0.9911026 | 0.9917382 | 0.9920559 | 0.9907849 | 0.9914204 | 0.9907849 | 0.9917382 | 0.9907849 | 0.9911026 |
4.00E-05 | 0.987925 | 0.9876072 | 0.9901493 | 0.9911026 | 0.9914204 | 0.9920559 | 0.9923737 | 0.9923737 | 0.9914204 | 0.9917382 | 0.9914204 | 0.9920559 | 0.9917382 | 0.9914204 | 0.9907849 | 0.9914204 | 0.9917382 | 0.9917382 | 0.9917382 | 0.9917382 | 0.9914204 | 0.9911026 | 0.9920559 | 0.9917382 | 0.9926915 |
3.00E-05 | 0.9885605 | 0.9860184 | 0.9898316 | 0.9917382 | 0.9923737 | 0.9936447 | 0.993327 | 0.9917382 | 0.9926915 | 0.9930092 | 0.9920559 | 0.993327 | 0.9923737 | 0.9920559 | 0.9926915 | 0.9920559 | 0.9917382 | 0.9930092 | 0.9920559 | 0.9911026 | 0.9920559 | 0.9917382 | 0.9926915 | 0.9920559 | 0.9917382 |
2.00E-05 | 0.9895138 | 0.9888783 | 0.9917382 | 0.9923737 | 0.9936447 | 0.9926915 | 0.9930092 | 0.9926915 | 0.9930092 | 0.9936447 | 0.9930092 | 0.9942803 | 0.9923737 | 0.9930092 | 0.993327 | 0.993327 | 0.9926915 | 0.9930092 | 0.993327 | 0.993327 | 0.9923737 | 0.9926915 | 0.9923737 | 0.9930092 | 0.9917382 |
1.00E-05 | 0.9904671 | 0.9888783 | 0.9917382 | 0.9930092 | 0.9930092 | 0.9930092 | 0.9939625 | 0.9936447 | 0.9942803 | 0.9930092 | 0.9942803 | 0.9936447 | 0.9939625 | 0.9939625 | 0.9930092 | 0.994598 | 0.9936447 | 0.9939625 | 0.9930092 | 0.9930092 | 0.9936447 | 0.9930092 | 0.993327 | 0.9936447 | 0.993327 |
网络的最大分类性能与平均性能变化的规律大体上是一致,随着卷积核数量的增加最大性能也是先上升后下降的规律。
但没有卷积核的裸网络的最大性能曲线仅略好于1个核的网络。裸网络的平均性能虽然与6个核的相当,但实验结果很清晰,只要卷积核的数量超过1个,获得更大分类准确率的概率就大幅增加。
3比较迭代次数
无核 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | 24 | |
迭代次数n | 迭代次数n | 迭代次数n | 迭代次数n | 迭代次数n | 迭代次数n | 迭代次数n | 迭代次数n | 迭代次数n | 迭代次数n | 迭代次数n | 迭代次数n | 迭代次数n | 迭代次数n | 迭代次数n | 迭代次数n | 迭代次数n | 迭代次数n | 迭代次数n | 迭代次数n | 迭代次数n | 迭代次数n | 迭代次数n | 迭代次数n | 迭代次数n | |
1.00E-04 | 28124.126 | 26363.337 | 29979.673 | 33505.824 | 34287.251 | 35781.005 | 36393 | 36198.372 | 36559.171 | 35908.97 | 35984.985 | 36432.995 | 35975.121 | 35158.608 | 34849.101 | 35285.01 | 35300.377 | 35492.181 | 34433.161 | 34965.156 | 34066.482 | 34509.965 | 33593.302 | 34826.045 | 33836.834 |
9.00E-05 | 29987.729 | 29978.573 | 32175.729 | 36229.457 | 36723.593 | 37055.271 | 38212.141 | 38496.814 | 38671.196 | 39193.96 | 38783.658 | 38386.97 | 38511.085 | 38309.382 | 38782.608 | 37330.683 | 37358.613 | 36983.894 | 37041.709 | 36984.116 | 36589.513 | 37149.889 | 36095.221 | 35700.322 | 35459.337 |
8.00E-05 | 31350.236 | 32322.02 | 34999.729 | 38706.91 | 40096.749 | 39660.045 | 41560.945 | 42218.603 | 40946.754 | 41336.392 | 41267.121 | 41394.698 | 40964.603 | 40719.739 | 39945.658 | 39421.854 | 40681.07 | 40038.734 | 38817.693 | 38490.392 | 38544.96 | 38975.573 | 38532.196 | 37816.598 | 37946.724 |
7.00E-05 | 34517.307 | 33146.799 | 37931.538 | 41524.055 | 43070.256 | 43073.96 | 44722.457 | 44130.141 | 44990.573 | 44166.347 | 43386.94 | 43004.307 | 43543.955 | 43540.221 | 42858.98 | 43018.231 | 42055.141 | 41833.317 | 41546.518 | 42022.985 | 40391.241 | 40626.05 | 41011.91 | 40349.578 | 40601.302 |
6.00E-05 | 38040.945 | 37338.347 | 41309.322 | 46602.578 | 46176.874 | 48387.608 | 48282.894 | 48211.799 | 47397.397 | 48940.141 | 48093.814 | 47818.92 | 47749.291 | 46612.362 | 46278.085 | 45673.824 | 45822.553 | 44635.226 | 44976.96 | 43909.548 | 44530.231 | 44611.427 | 43462.874 | 44263.307 | 43895.166 |
5.00E-05 | 44035.538 | 41032.764 | 46641.221 | 50103.844 | 51737.774 | 54111.623 | 53573.492 | 53435.427 | 53356.799 | 53122.553 | 51852.688 | 52563.638 | 52129.608 | 51287.889 | 51800.347 | 49121.633 | 48671.849 | 49192.894 | 48718.191 | 49348.744 | 47596.407 | 48566.231 | 48409.462 | 48112.05 | 48021.271 |
4.00E-05 | 51206.688 | 46429.281 | 51540.754 | 56576.111 | 59430.935 | 58390.814 | 60279.035 | 61504.673 | 59402.156 | 58477.593 | 57556.658 | 58451.106 | 58198 | 56739.638 | 55744.869 | 55820.08 | 55373.809 | 55034.271 | 55470.683 | 54251.538 | 54618.156 | 53051.784 | 53374.99 | 52447.206 | 53108.342 |
3.00E-05 | 60094.633 | 53295.553 | 60681.412 | 65860.221 | 67512.538 | 69861.226 | 69671.769 | 68288.568 | 68381.683 | 67892.543 | 65794.302 | 66697.769 | 66502.03 | 65532.558 | 65526.543 | 64026.447 | 62647 | 63022.111 | 63046.719 | 60707.874 | 60312.698 | 59228.307 | 61837.025 | 59463.372 | 59193.749 |
2.00E-05 | 75255.03 | 68115.648 | 73774.844 | 78836.91 | 82261.482 | 84607.482 | 82296.538 | 82697.447 | 81852.322 | 83528.045 | 80141.94 | 79658.09 | 77541.94 | 76919.688 | 77456.201 | 77743.704 | 76069.894 | 74429.814 | 75165.925 | 72190.055 | 71793.789 | 72919.648 | 71729.116 | 72143.603 | 69703.226 |
1.00E-05 | 105617.92 | 89846.487 | 98724.668 | 108704.93 | 108726.58 | 113518.73 | 111453.34 | 108518.19 | 108211.67 | 108714.66 | 107922.2 | 104651.31 | 104601.31 | 103637.22 | 101395.85 | 102920.23 | 101257.73 | 102736.07 | 100495.92 | 97050.97 | 98032.246 | 97393.844 | 94972.337 | 95764.648 | 95546.337 |
迭代次数曲线的变化规律是最为清晰的,迭代次数随着卷积核数量的增加是先增加后下降的,无卷积核的裸网络的迭代次数曲线和2个核的迭代次数曲线大体重合。
迭代次数曲线的峰值出现在5个核。这个位置比平均性能曲线的峰值的位置稍小些。
1<2=无核<3<4<5>6>7>8>9>10>11>12>13>14>15>16>17>18>19>20>21>22>23>24
最后考虑耗时
无核 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | 24 | |
耗时 min/198 | 耗时 min/199 | 耗时 min/199 | 耗时 min/199 | 耗时 min/199 | 耗时 min/199 | 耗时 min/199 | 耗时 min/199 | 耗时 min/199 | 耗时 min/199 | 耗时 min/199 | 耗时 min/199 | 耗时 min/199 | 耗时 min/199 | 耗时 min/199 | 耗时 min/199 | 耗时 min/199 | 耗时 min/199 | 耗时 min/199 | 耗时 min/199 | 耗时 min/199 | 耗时 min/199 | 耗时 min/199 | 耗时 min/199 | 耗时 min/199 | |
1.00E-04 | 1.7959833 | 5.2044667 | 10.761983 | 17.697133 | 25.06455 | 30.761717 | 38.807083 | 43.825117 | 51.394867 | 55.337333 | 62.993983 | 69.109817 | 76.503917 | 79.68595 | 85.922717 | 93.397983 | 99.499967 | 107.68242 | 112.55437 | 113.82033 | 118.0696 | 123.97782 | 127.19317 | 142.86933 | 141.31773 |
9.00E-05 | 1.9285167 | 5.9544 | 11.538767 | 19.178217 | 24.9489 | 33.310783 | 41.97085 | 47.8454 | 54.231983 | 63.54735 | 67.782983 | 72.293983 | 79.248883 | 89.167317 | 95.01835 | 98.188633 | 105.54232 | 111.51525 | 117.93158 | 122.85382 | 124.33822 | 130.09507 | 137.69198 | 140.9868 | 147.36062 |
8.00E-05 | 2.0181 | 6.4226 | 12.526817 | 20.4336 | 28.973583 | 34.233083 | 43.31305 | 51.449867 | 57.095767 | 62.64085 | 72.13595 | 78.326667 | 87.38795 | 93.061933 | 98.24765 | 103.96745 | 113.85245 | 118.3822 | 124.19218 | 125.0228 | 131.31508 | 137.47255 | 145.24483 | 150.58187 | 156.24917 |
7.00E-05 | 2.2135 | 6.57305 | 13.5572 | 22.3224 | 28.419017 | 37.157617 | 48.042383 | 54.732367 | 62.63395 | 69.25505 | 75.80715 | 82.404633 | 90.20145 | 100.91835 | 105.26863 | 113.65118 | 118.32975 | 124.60877 | 131.0534 | 137.19933 | 138.18822 | 143.59333 | 154.67182 | 165.37337 | 167.5772 |
6.00E-05 | 2.43035 | 7.3859 | 14.74005 | 24.50945 | 32.6907 | 42.592317 | 52.715617 | 57.891017 | 66.566933 | 76.920033 | 83.9125 | 88.9966 | 100.0437 | 106.06915 | 113.02703 | 117.8712 | 128.56613 | 133.72853 | 144.64843 | 144.04102 | 151.59498 | 164.9672 | 163.39712 | 175.59025 | 180.9757 |
5.00E-05 | 2.80525 | 8.16315 | 15.777617 | 25.080683 | 37.016817 | 47.963317 | 56.428817 | 64.824583 | 73.47375 | 83.822367 | 90.14335 | 98.567317 | 110.75688 | 115.40903 | 126.26177 | 126.53413 | 135.83447 | 147.07502 | 154.37767 | 161.01043 | 160.36768 | 177.78613 | 183.40363 | 191.68515 | 197.76138 |
4.00E-05 | 3.2485167 | 7.1923667 | 18.822567 | 29.470433 | 43.93925 | 50.1656 | 63.04765 | 74.211633 | 83.287767 | 91.3095 | 99.89205 | 108.23228 | 121.45302 | 128.2457 | 136.08832 | 144.60355 | 154.1469 | 163.62747 | 175.97142 | 176.54183 | 190.67043 | 194.9548 | 202.14387 | 208.87257 | 216.197 |
3.00E-05 | 3.8048 | 10.5643 | 22.00195 | 34.03475 | 44.198017 | 59.995433 | 72.538433 | 82.9001 | 94.604783 | 105.82542 | 113.39317 | 124.97498 | 139.03407 | 147.11228 | 159.68545 | 164.49283 | 174.89538 | 188.46768 | 193.85952 | 194.90477 | 205.81675 | 216.59527 | 232.56082 | 233.77435 | 242.31962 |
2.00E-05 | 4.62085 | 13.376017 | 24.850633 | 40.5423 | 57.55195 | 72.333867 | 87.097033 | 99.13375 | 113.02132 | 130.37197 | 138.93015 | 148.53678 | 164.89943 | 173.81777 | 187.39533 | 200.17655 | 209.38253 | 220.41147 | 229.60457 | 233.57257 | 246.46873 | 255.69252 | 270.11763 | 283.36797 | 285.86822 |
1.00E-05 | 6.55745 | 15.9231 | 35.727033 | 55.903167 | 75.809167 | 96.261717 | 114.3315 | 130.69715 | 150.03637 | 167.87448 | 184.22573 | 196.991 | 217.31047 | 230.28493 | 244.44427 | 265.8767 | 278.56902 | 303.39292 | 307.10363 | 311.7583 | 327.96728 | 339.95648 | 354.93202 | 373.60505 | 389.59185 |
这条曲线接近直线,收敛标准越小直线的斜率越大。无核的裸网络的的速度是最快的,无核的裸网络的平均性能相当于6个核,但收敛之间只有6个核的5.7%。
8 | 24 | |
1.00E-05 | 1.00E-05 | |
150.03637 | 389.5919 | 耗时 min/199 |
0.9893669 | 0.988632 | 平均准确率p-ave |
0.9992576 | 2.596649 |
24核的网络的性能只有8核性能的99.92%,损失了0.07.5%,但是耗时是8核的2.59倍。
当δ=1e-5时,耗时y和卷积核数量x之间有关系
Y=15.897015275260872*x+16.644531281739127
0.9936006780703771 ****** 决定系数 r**2
可以用此推算当卷积核的数量是100个的时候耗时为1606.3460588078262min,是8核的10.7倍,性能损失至少大于0.075%,也就是当有100个核时的平均性能比较峰值性能损失至少0.075%,但要多花10倍的时间。由此可见一个不恰当的卷积核数量设置可能会导致巨大的算力浪费。
无核<1<2<3<4<5<6<7<8<9<10<11<12<13<14<15<16<17<18<19<20<21<22<23<24
用迭代次数的数据/耗时得到:次/min的数据
无核 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | 24 | |
次/min | 次/min | 次/min | 次/min | 次/min | 次/min | 次/min | 次/min | 次/min | 次/min | 次/min | 次/min | 次/min | 次/min | 次/min | 次/min | 次/min | 次/min | 次/min | 次/min | 次/min | 次/min | 次/min | 次/min | 次/min | |
1.00E-04 | 15659.46 | 5065.521 | 2785.702 | 1893.291 | 1367.958 | 1163.166718 | 937.7927 | 825.9732 | 711.33896 | 648.9104 | 571.2447923 | 527.1753961 | 470.2389 | 441.2146 | 405.5865766 | 377.792 | 354.7778 | 329.6005 | 305.9247 | 307.196 | 288.5288 | 278.356 | 264.1125 | 243.7615 | 239.4379911 |
9.00E-05 | 15549.63 | 5034.692 | 2788.489 | 1889.094 | 1471.952 | 1112.410687 | 910.4448 | 804.6085 | 713.06992 | 616.7678 | 572.1739644 | 530.9842961 | 485.9511 | 429.6348 | 408.1591402 | 380.1935 | 353.9681 | 331.6488 | 314.0949 | 301.0416 | 294.2741 | 285.5596 | 262.1447 | 253.2175 | 240.6296708 |
8.00E-05 | 15534.53 | 5032.544 | 2793.984 | 1894.278 | 1383.907 | 1158.529743 | 959.5479 | 820.5775 | 717.15919 | 659.8951 | 572.0742654 | 528.4879372 | 468.7672 | 437.5553 | 406.5813105 | 379.175 | 357.314 | 338.2158 | 312.5615 | 307.867 | 293.5303 | 283.5153 | 265.2913 | 251.1365 | 242.8603264 |
7.00E-05 | 15593.99 | 5042.834 | 2797.889 | 1860.197 | 1515.544 | 1159.222891 | 930.8959 | 806.2896 | 718.30968 | 637.7347 | 572.3330807 | 521.8675794 | 482.7412 | 431.4401 | 407.139131 | 378.5111 | 355.4063 | 335.7173 | 317.0198 | 306.2915 | 292.2915 | 282.9243 | 265.1544 | 243.9908 | 242.2841622 |
6.00E-05 | 15652.46 | 5055.355 | 2802.522 | 1901.413 | 1412.539 | 1136.064244 | 915.9125 | 832.8028 | 712.02615 | 636.247 | 573.142429 | 537.311758 | 477.2843 | 439.4526 | 409.4426268 | 387.4893 | 356.4123 | 333.7749 | 310.9398 | 304.8406 | 293.7448 | 270.426 | 265.9954 | 252.0829 | 242.5472913 |
5.00E-05 | 15697.54 | 5026.585 | 2956.164 | 1997.707 | 1397.683 | 1128.187683 | 949.3995 | 824.3081 | 726.2022 | 633.7515 | 575.2247774 | 533.2765461 | 470.667 | 444.401 | 410.2615392 | 388.2086 | 358.3174 | 334.4749 | 315.578 | 306.4941 | 296.7955 | 273.1722 | 263.9504 | 250.9952 | 242.8243095 |
4.00E-05 | 15763.1 | 6455.355 | 2738.243 | 1919.758 | 1352.571 | 1163.961242 | 956.0869 | 828.774 | 713.21586 | 640.4327 | 576.1885785 | 540.0524107 | 479.1812 | 442.4292 | 409.6227414 | 386.0215 | 359.2275 | 336.3388 | 315.2255 | 307.3013 | 286.4532 | 272.1235 | 264.0446 | 251.0967 | 245.6479124 |
3.00E-05 | 15794.43 | 5044.873 | 2758.002 | 1935.088 | 1527.502 | 1164.442396 | 960.4807 | 823.7453 | 722.81423 | 641.5523 | 580.2316263 | 533.6889597 | 478.3146 | 445.4595 | 410.3476097 | 389.2355 | 358.197 | 334.3921 | 325.2186 | 311.4745 | 293.0408 | 273.4515 | 265.8961 | 254.3623 | 244.2796401 |
2.00E-05 | 16285.97 | 5092.372 | 2968.731 | 1944.559 | 1429.343 | 1169.680073 | 944.8834 | 834.2007 | 724.22021 | 640.6902 | 576.8505951 | 536.2852801 | 470.2378 | 442.5306 | 413.3304689 | 388.3757 | 363.3058 | 337.6858 | 327.3712 | 309.0691 | 291.2896 | 285.1849 | 265.5477 | 254.5934 | 243.8299261 |
1.00E-05 | 16106.55 | 5642.525 | 2763.304 | 1944.522 | 1434.214 | 1179.27186 | 974.8262 | 830.3026 | 721.23626 | 647.5949 | 585.8150165 | 531.2491765 | 481.345 | 450.0391 | 414.8015031 | 387.0976 | 363.4924 | 338.6238 | 327.2378 | 311.302 | 298.9086 | 286.4892 | 267.579 | 256.3259 | 245.2472676 |
将速度曲线画成图,拟合
y=5630.384286483394*x**-0.9849340360486022
0.9997146318921649 ****** 决定系数 r**2
X是卷积核的数量单位:个
Y是网络迭代速度单位:次/分钟
综合起来看,平均性能的峰值出现在8核,迭代次数的峰值出现在5核,
5 | 8 | |
1.00E-05 | 1.00E-05 | |
96.2617167 | 150.036367 | 耗时 min/199 |
0.98900604 | 0.98936692 | 平均准确率p-ave |
1.00036489 | 1.55862966 |
对比8核和5核的数据,8核的平均性能要好0.03%,但耗时是5核的1.55倍,因此从收敛效率考虑卷积核数量等于5是一个综合性能比较全面的设置方案。