构造一个81*30*3的网络和一个有5个3*3卷积核的网络,分类mnist的0,1,2通过分类准确率数据比较卷积核的作用。
数据集是mnist的0,1,2将图片变成9*9,
无卷积核的裸网络的结构是
Mnist(0,1,2)---81---30---3----(1,0,0) || (0,1,0) || (0,0,1)
5个卷积核的网络的结构是
Mnist(0,1,2)---con3*3*5---7*7*5---30---3----(1,0,0) || (0,1,0) || (0,0,1)
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-4到1e-8.每个δ收敛199次取平均值,共收敛了2*17*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和199测分类准确率的标准差。
无核网路的数据
无核 | ||||||||||
f2[0] | f2[1] | f2[2] | 迭代次数n | 平均准确率p-ave | δ | 耗时ms/次 | 耗时ms/199次 | 耗时 min/199 | 最大值p-max | 标准差 |
7.44E-05 | 0.196045 | 0.803986 | 27833.26 | 0.982809 | 1.00E-04 | 547.1005 | 108896 | 1.814933 | 0.986336 | 0.00211 |
6.83E-05 | 0.175937 | 0.824082 | 29340.49 | 0.982986 | 9.00E-05 | 573.1156 | 114054 | 1.9009 | 0.986336 | 0.00188 |
5.85E-05 | 0.155833 | 0.844184 | 31737.18 | 0.983441 | 8.00E-05 | 618.3266 | 123052 | 2.050867 | 0.986336 | 0.001577 |
5.28E-05 | 0.140752 | 0.859263 | 35173.1 | 0.983954 | 7.00E-05 | 683.9799 | 136117 | 2.268617 | 0.986972 | 0.00156 |
4.35E-05 | 0.125671 | 0.874342 | 39840.33 | 0.984409 | 6.00E-05 | 768.4221 | 152917 | 2.548617 | 0.986972 | 0.001206 |
3.62E-05 | 0.145763 | 0.854247 | 44351.31 | 0.984731 | 5.00E-05 | 852.5075 | 169655 | 2.827583 | 0.987925 | 0.001112 |
2.74E-05 | 0.020133 | 0.979874 | 50897.51 | 0.985356 | 4.00E-05 | 983.9598 | 195809 | 3.263483 | 0.987607 | 0.000962 |
2.08E-05 | 0.050275 | 0.949732 | 58540.31 | 0.985867 | 3.00E-05 | 1116.874 | 222260 | 3.704333 | 0.988561 | 0.000888 |
1.38E-05 | 0.025142 | 0.974863 | 75020.01 | 0.986892 | 2.00E-05 | 1426.417 | 283873 | 4.731217 | 0.989514 | 0.000979 |
6.59E-06 | 0.030159 | 0.969844 | 105302.8 | 0.988173 | 1.00E-05 | 1998.704 | 397760 | 6.629333 | 0.990467 | 0.000959 |
5.49E-06 | 0.025133 | 0.97487 | 112076.2 | 0.988353 | 9.00E-06 | 2110.879 | 420065 | 7.001083 | 0.990149 | 0.000943 |
5.04E-06 | 0.010057 | 0.989946 | 117964.6 | 0.988484 | 8.00E-06 | 1685.663 | 335447 | 5.590783 | 0.990149 | 0.000934 |
4.40E-06 | 0.005031 | 0.994971 | 121447.5 | 0.988562 | 7.00E-06 | 2273.709 | 452468 | 7.541133 | 0.990467 | 0.000869 |
3.57E-06 | 0.00503 | 0.994972 | 131816.6 | 0.988819 | 6.00E-06 | 2486.296 | 494773 | 8.246217 | 0.990467 | 0.000776 |
3.06E-06 | 0.00503 | 0.994972 | 139246.6 | 0.988963 | 5.00E-06 | 2608.704 | 519148 | 8.652467 | 0.990785 | 0.000846 |
2.38E-06 | 3.57E-06 | 0.999998 | 151750.2 | 0.989309 | 4.00E-06 | 2832.452 | 563658 | 9.3943 | 0.991103 | 0.000704 |
1.69E-06 | 2.61E-06 | 0.999998 | 171937 | 0.989515 | 3.00E-06 | 3367.844 | 670218 | 11.1703 | 0.991738 | 0.00071 |
1.17E-06 | 1.70E-06 | 0.999999 | 192145.3 | 0.989838 | 2.00E-06 | 3892.382 | 774592 | 12.90987 | 0.992374 | 0.000737 |
5.27E-07 | 8.49E-07 | 0.999999 | 239406.2 | 0.990485 | 1.00E-06 | 4200.975 | 836004 | 13.9334 | 0.992056 | 0.00071 |
4.82E-07 | 7.58E-07 | 0.999999 | 242219.7 | 0.990493 | 9.00E-07 | 4819.492 | 959087 | 15.98478 | 0.992056 | 0.000757 |
4.10E-07 | 6.94E-07 | 1 | 252592.7 | 0.990489 | 8.00E-07 | 4682.965 | 931916 | 15.53193 | 0.992374 | 0.000747 |
3.61E-07 | 6.05E-07 | 1 | 264546.9 | 0.990627 | 7.00E-07 | 4006.704 | 797340 | 13.289 | 0.992374 | 0.000802 |
3.01E-07 | 5.15E-07 | 1 | 270883.3 | 0.990676 | 6.00E-07 | 4904.412 | 975989 | 16.26648 | 0.992691 | 0.000774 |
2.40E-07 | 4.36E-07 | 1 | 288360.4 | 0.990876 | 5.00E-07 | 5213.447 | 1037490 | 17.2915 | 0.992374 | 0.000763 |
1.96E-07 | 3.47E-07 | 1 | 298620.9 | 0.990921 | 4.00E-07 | 5504.915 | 1095486 | 18.2581 | 0.992691 | 0.000761 |
1.40E-07 | 2.64E-07 | 1 | 321430.4 | 0.991157 | 3.00E-07 | 5984.598 | 1190954 | 19.84923 | 0.992691 | 0.000749 |
9.61E-08 | 1.74E-07 | 1 | 352910.5 | 0.991388 | 2.00E-07 | 6380.844 | 1269793 | 21.16322 | 0.993327 | 0.00081 |
4.34E-08 | 8.83E-08 | 1 | 400271.1 | 0.991641 | 1.00E-07 | 7226.673 | 1438111 | 23.96852 | 0.993327 | 0.000805 |
3.55E-08 | 8.02E-08 | 1 | 418117.3 | 0.991618 | 9.00E-08 | 7554.206 | 1503293 | 25.05488 | 0.993327 | 0.000814 |
3.17E-08 | 7.15E-08 | 1 | 422921 | 0.991623 | 8.00E-08 | 7639.899 | 1520351 | 25.33918 | 0.993645 | 0.000817 |
2.78E-08 | 6.30E-08 | 1 | 431914.3 | 0.991701 | 7.00E-08 | 7826.965 | 1557578 | 25.95963 | 0.993963 | 0.000806 |
2.50E-08 | 5.31E-08 | 1 | 450995.3 | 0.99169 | 6.00E-08 | 8424.93 | 1676577 | 27.94295 | 0.993327 | 0.000812 |
1.92E-08 | 4.46E-08 | 1 | 451507.8 | 0.991741 | 5.00E-08 | 7867.839 | 1565708 | 26.09513 | 0.993327 | 0.000764 |
1.51E-08 | 3.63E-08 | 1 | 477393.6 | 0.991709 | 4.00E-08 | 8821.698 | 1755519 | 29.25865 | 0.993645 | 0.000898 |
1.07E-08 | 2.68E-08 | 1 | 500028.4 | 0.991799 | 3.00E-08 | 8999.995 | 1791000 | 29.85 | 0.993645 | 0.000833 |
7.50E-09 | 1.79E-08 | 1 | 525958.1 | 0.991767 | 2.00E-08 | 10181.58 | 2026156 | 33.76927 | 0.99428 | 0.000815 |
4.07E-09 | 8.80E-09 | 1 | 576699 | 0.991812 | 1.00E-08 | 10419.82 | 2073544 | 34.55907 | 0.993963 | 0.000785 |
5个卷积核网络的数据
5 | ||||||||||
f2[0] | f2[1] | f2[2] | 迭代次数n | 平均准确率p-ave | δ | 耗时ms/次 | 耗时ms/199次 | 耗时 min/199 | 最大值p-max | 标准差 |
0.698479 | 0.286469 | 0.015158 | 35781.01 | 0.98182 | 1.00E-04 | 9274.779 | 1845703 | 30.76172 | 0.990149 | 0.005496 |
0.71355 | 0.251294 | 0.035248 | 37055.27 | 0.98186 | 9.00E-05 | 10043.37 | 1998647 | 33.31078 | 0.990785 | 0.005262 |
0.743701 | 0.231193 | 0.025191 | 39660.05 | 0.98263 | 8.00E-05 | 10321.43 | 2053985 | 34.23308 | 0.991103 | 0.005188 |
0.723605 | 0.251285 | 0.025183 | 43073.96 | 0.983444 | 7.00E-05 | 11203.3 | 2229457 | 37.15762 | 0.99142 | 0.004914 |
0.743706 | 0.226156 | 0.030199 | 48387.61 | 0.984043 | 6.00E-05 | 12841.87 | 2555539 | 42.59232 | 0.990785 | 0.00463 |
0.788931 | 0.165856 | 0.045265 | 54111.62 | 0.984663 | 5.00E-05 | 14461.2 | 2877799 | 47.96332 | 0.991738 | 0.003962 |
0.839182 | 0.110578 | 0.050282 | 58390.81 | 0.984872 | 4.00E-05 | 15125.22 | 3009936 | 50.1656 | 0.992056 | 0.004193 |
0.819085 | 0.125647 | 0.055299 | 69861.23 | 0.986394 | 3.00E-05 | 18088.98 | 3599726 | 59.99543 | 0.993645 | 0.003836 |
0.934664 | 0.025141 | 0.040216 | 84607.48 | 0.987275 | 2.00E-05 | 21809.21 | 4340032 | 72.33387 | 0.992691 | 0.003321 |
0.884417 | 0.040209 | 0.075384 | 113518.7 | 0.989006 | 1.00E-05 | 29023.62 | 5775703 | 96.26172 | 0.993009 | 0.002595 |
0.884418 | 0.035183 | 0.080408 | 116764.5 | 0.989011 | 9.00E-06 | 31420.88 | 6252772 | 104.2129 | 0.993645 | 0.002452 |
0.924619 | 0.030157 | 0.045232 | 121199.8 | 0.988969 | 8.00E-06 | 32736.33 | 6514563 | 108.5761 | 0.993963 | 0.002788 |
0.864318 | 0.015081 | 0.120607 | 127616.1 | 0.989206 | 7.00E-06 | 33793.67 | 6724962 | 112.0827 | 0.993009 | 0.002299 |
0.874369 | 0.035181 | 0.090456 | 133287.6 | 0.989308 | 6.00E-06 | 35072.14 | 6979369 | 116.3228 | 0.993963 | 0.002649 |
0.844219 | 0.03518 | 0.120606 | 148576.1 | 0.989966 | 5.00E-06 | 38453.66 | 7652282 | 127.538 | 0.99428 | 0.002038 |
0.909546 | 0.010054 | 0.080405 | 153811.6 | 0.990119 | 4.00E-06 | 40236.49 | 8007064 | 133.4511 | 0.993645 | 0.001917 |
0.783919 | 0.020103 | 0.195981 | 174867 | 0.989859 | 3.00E-06 | 45900.82 | 9134269 | 152.2378 | 0.99428 | 0.002644 |
0.778894 | 0.020102 | 0.201006 | 191543.2 | 0.990271 | 2.00E-06 | 50249.96 | 9999758 | 166.6626 | 0.993963 | 0.002102 |
0.758794 | 0.010051 | 0.231156 | 230501 | 0.990403 | 1.00E-06 | 59820.19 | 11904219 | 198.4037 | 0.993963 | 0.002415 |
0.653266 | 0.015076 | 0.331659 | 240200.4 | 0.990157 | 9.00E-07 | 62553.32 | 12448114 | 207.4686 | 0.994598 | 0.00262 |
0.663317 | 0.010051 | 0.326633 | 248517.1 | 0.990466 | 8.00E-07 | 64427.78 | 12821141 | 213.6857 | 0.99428 | 0.002311 |
0.678392 | 0.020101 | 0.301508 | 257205 | 0.990272 | 7.00E-07 | 68003.96 | 13532795 | 225.5466 | 0.993963 | 0.002239 |
0.723618 | 0.015076 | 0.261307 | 263480.2 | 0.990697 | 6.00E-07 | 67587.88 | 13450013 | 224.1669 | 0.995234 | 0.002069 |
0.61809 | 4.36E-07 | 0.38191 | 275964 | 0.99074 | 5.00E-07 | 71423.62 | 14213308 | 236.8885 | 0.994916 | 0.002054 |
0.59799 | 0.025126 | 0.376885 | 293300.9 | 0.990467 | 4.00E-07 | 75939.62 | 15111996 | 251.8666 | 0.99428 | 0.002173 |
0.562814 | 2.61E-07 | 0.437186 | 310972.9 | 0.990822 | 3.00E-07 | 80267.45 | 15973228 | 266.2205 | 0.994916 | 0.002226 |
0.547739 | 0.01005 | 0.442211 | 336482.8 | 0.990976 | 2.00E-07 | 87635.89 | 17439546 | 290.6591 | 0.995234 | 0.002104 |
0.532663 | 0.015075 | 0.452261 | 382286.8 | 0.990991 | 1.00E-07 | 99956.42 | 19891345 | 331.5224 | 0.994598 | 0.002179 |
0.492462 | 0.005025 | 0.502513 | 388770.6 | 0.990823 | 9.00E-08 | 101827 | 20263594 | 337.7266 | 0.994916 | 0.002112 |
0.462312 | 0.005025 | 0.532663 | 390518.9 | 0.990989 | 8.00E-08 | 102264.8 | 20350700 | 339.1783 | 0.994598 | 0.002224 |
0.452261 | 0.005025 | 0.542714 | 400328.2 | 0.990715 | 7.00E-08 | 104427.7 | 20781121 | 346.352 | 0.994916 | 0.002378 |
0.40201 | 0.015075 | 0.582915 | 423798.3 | 0.991336 | 6.00E-08 | 111411.3 | 22170863 | 369.5144 | 0.995234 | 0.002112 |
0.417085 | 0.015075 | 0.567839 | 437013.5 | 0.991344 | 5.00E-08 | 113981.6 | 22682349 | 378.0392 | 0.994916 | 0.002024 |
0.407035 | 0.01005 | 0.582915 | 459736.3 | 0.991328 | 4.00E-08 | 120307.9 | 23941283 | 399.0214 | 0.994598 | 0.001805 |
0.407035 | 0.005025 | 0.58794 | 470944.5 | 0.991096 | 3.00E-08 | 122076.8 | 24293292 | 404.8882 | 0.995869 | 0.001853 |
0.356784 | 0.015075 | 0.628141 | 502589.2 | 0.991781 | 2.00E-08 | 131185.6 | 26105935 | 435.0989 | 0.994916 | 0.001486 |
0.346734 | 0.005025 | 0.648241 | 572687 | 0.991658 | 1.00E-08 | 150921.6 | 30033399 | 500.5567 | 0.994598 | 0.001607 |
将平均准确率和最大准确率以及标准差列出来
无核 | 5个核 | 无核 | 5个核 | 无核 | 5个核 | ||||
δ | 平均准确率p-ave | 平均准确率p-ave | 最大值p-max | 最大值p-max | 标准差 | 标准差 | |||
1.00E-04 | 0.982809 | 0.98182 | 0.986336 | 0.990149 | 0.00211 | 0.005496 | 0.383995 | ||
9.00E-05 | 0.982986 | 0.98186 | 0.986336 | 0.990785 | 0.00188 | 0.005262 | 0.357206 | ||
8.00E-05 | 0.983441 | 0.98263 | 0.986336 | 0.991103 | 0.001577 | 0.005188 | 0.303999 | ||
7.00E-05 | 0.983954 | 0.983444 | 0.986972 | 0.99142 | 0.00156 | 0.004914 | 0.31742 | ||
6.00E-05 | 0.984409 | 0.984043 | 0.986972 | 0.990785 | 0.001206 | 0.00463 | 0.260549 | ||
5.00E-05 | 0.984731 | 0.984663 | 0.987925 | 0.991738 | 0.001112 | 0.003962 | 0.280694 | ||
4.00E-05 | 0.985356 | 0.984872 | 0.987607 | 0.992056 | 0.000962 | 0.004193 | 0.229515 | ||
3.00E-05 | 0.985867 | 0.986394 | 0.988561 | 0.993645 | 0.000888 | 0.003836 | 0.23158 | ||
2.00E-05 | 0.986892 | 0.987275 | 0.989514 | 0.992691 | 0.000979 | 0.003321 | 0.294673 | ||
1.00E-05 | 0.988173 | 0.989006 | 0.990467 | 0.993009 | 0.000959 | 0.002595 | 0.369499 | ||
9.00E-06 | 0.988353 | 0.989011 | 0.990149 | 0.993645 | 0.000943 | 0.002452 | 0.384622 | ||
8.00E-06 | 0.988484 | 0.988969 | 0.990149 | 0.993963 | 0.000934 | 0.002788 | 0.334908 | ||
7.00E-06 | 0.988562 | 0.989206 | 0.990467 | 0.993009 | 0.000869 | 0.002299 | 0.378022 | ||
6.00E-06 | 0.988819 | 0.989308 | 0.990467 | 0.993963 | 0.000776 | 0.002649 | 0.293088 | ||
5.00E-06 | 0.988963 | 0.989966 | 0.990785 | 0.99428 | 0.000846 | 0.002038 | 0.41509 | ||
4.00E-06 | 0.989309 | 0.990119 | 0.991103 | 0.993645 | 0.000704 | 0.001917 | 0.367068 | ||
3.00E-06 | 0.989515 | 0.989859 | 0.991738 | 0.99428 | 0.00071 | 0.002644 | 0.268445 | ||
2.00E-06 | 0.989838 | 0.990271 | 0.992374 | 0.993963 | 0.000737 | 0.002102 | 0.350831 | ||
1.00E-06 | 0.990485 | 0.990403 | 0.992056 | 0.993963 | 0.00071 | 0.002415 | 0.293812 | ||
9.00E-07 | 0.990493 | 0.990157 | 0.992056 | 0.994598 | 0.000757 | 0.00262 | 0.28912 | ||
8.00E-07 | 0.990489 | 0.990466 | 0.992374 | 0.99428 | 0.000747 | 0.002311 | 0.323109 | ||
7.00E-07 | 0.990627 | 0.990272 | 0.992374 | 0.993963 | 0.000802 | 0.002239 | 0.358189 | ||
6.00E-07 | 0.990676 | 0.990697 | 0.992691 | 0.995234 | 0.000774 | 0.002069 | 0.374171 | ||
5.00E-07 | 0.990876 | 0.99074 | 0.992374 | 0.994916 | 0.000763 | 0.002054 | 0.371331 | ||
4.00E-07 | 0.990921 | 0.990467 | 0.992691 | 0.99428 | 0.000761 | 0.002173 | 0.35034 | ||
3.00E-07 | 0.991157 | 0.990822 | 0.992691 | 0.994916 | 0.000749 | 0.002226 | 0.336619 | ||
2.00E-07 | 0.991388 | 0.990976 | 0.993327 | 0.995234 | 0.00081 | 0.002104 | 0.384796 | ||
1.00E-07 | 0.991641 | 0.990991 | 0.993327 | 0.994598 | 0.000805 | 0.002179 | 0.369567 | ||
9.00E-08 | 0.991618 | 0.990823 | 0.993327 | 0.994916 | 0.000814 | 0.002112 | 0.385203 | ||
8.00E-08 | 0.991623 | 0.990989 | 0.993645 | 0.994598 | 0.000817 | 0.002224 | 0.367109 | ||
7.00E-08 | 0.991701 | 0.990715 | 0.993963 | 0.994916 | 0.000806 | 0.002378 | 0.338951 | ||
6.00E-08 | 0.99169 | 0.991336 | 0.993327 | 0.995234 | 0.000812 | 0.002112 | 0.384475 | ||
5.00E-08 | 0.991741 | 0.991344 | 0.993327 | 0.994916 | 0.000764 | 0.002024 | 0.377483 | ||
4.00E-08 | 0.991709 | 0.991328 | 0.993645 | 0.994598 | 0.000898 | 0.001805 | 0.497674 | ||
3.00E-08 | 0.991799 | 0.991096 | 0.993645 | 0.995869 | 0.000833 | 0.001853 | 0.449515 | ||
2.00E-08 | 0.991767 | 0.991781 | 0.99428 | 0.994916 | 0.000815 | 0.001486 | 0.548764 | ||
1.00E-08 | 0.991812 | 0.991658 | 0.993963 | 0.994598 | 0.000785 | 0.001607 | 0.488578 |
观察平均准确率
可以看到无核的裸网络和含卷积核的网络的平均性能其实相差不大,
在δ>1e-6的区间含卷积核的网络的平均性能>无核的裸网络;
在δ<1e-6的区间无核的裸网络的平均性能>含卷积核的网络;
第二步比较标准差
这个很明显,网络增加了卷积核只是使网络平均性能的标准差变大了,也就是无核的裸网络的性能要比增加了卷积核的网络要稳定。裸网络的标准差只有含卷积核网络的1/3左右。
最后比较最大性能
因为裸网络和含有卷积核的网络的平均性能差不多,但是含卷积核网路的标准差更大显然使得含卷积核网络的最大性能也更大。
因此,卷积核只是增加了网络的性能的标准差,让网络变得更不稳定,增加了得到更大准确率的概率当然也增加了得到更差准确率的概率。但就平均性能而言和无核的裸网络平均性能相差不大,但却慢的多。
至少对这个实验而言一个结构简单的网络的性能要远好于含有多个卷积核的复杂网络。
***这篇文的实验数据已经上传到我的资源