是否可能通过无限增加卷积核的办法使网络性能无限提升?
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。