用7*7的卷积核分类9*9的图片到底应该用几个卷积核?55个

(mnist 0,2)-con(7*7)*n-30*2-(1,0)(0,1)

用7*7的卷积核分类mnist的0和2，将图片用间隔取点的办法缩小到9*9。从1开始不断增加卷积核的数量，观察网络的性能是如何随着卷积核的数量的变化而变化的。卷积核的数量从1到80个，收敛标准δ=6e-5,每个收敛标准收敛199次，统计平均值。

得到的表格

卷积核数量	f2[0]	f2[1]	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
0	5.12E-05	0.999949	3902.291	0.981170764	6E-05	88.13065	17538	0.2923	0.98508946
1	0.070397	0.929603	20632.15	0.976048233	6E-05	3830.975	762364	12.70607	0.98807157
2	0.055323	0.944677	10066.22	0.981190745	6E-05	4064.412	808821	13.48035	0.99005964
3	0.045275	0.954725	9838.578	0.983136358	6E-05	5983.166	1190666	19.84443	0.9915507
4	0.04025	0.959749	10507.51	0.984260268	6E-05	8212.638	1634323	27.23872	0.99005964
5	0.045275	0.954724	11317.55	0.984794749	6E-05	10963.5	2181743	36.36238	0.99055666
6	0.0503	0.9497	12220.42	0.986063518	6E-05	14357.38	2857119	47.61865	0.99204771
7	0.0503	0.9497	12511.11	0.986085997	6E-05	17324.5	3447606	57.4601	0.99105368
8	0.055325	0.944675	12953.54	0.986218368	6E-05	19259.08	3832562	63.87603	0.99055666
9	0.060349	0.939651	13212.72	0.986083499	6E-05	22801.35	4537477	75.62462	0.9915507
10	0.070399	0.929601	13858.86	0.985991089	6E-05	26756.56	5324559	88.74265	0.99105368
11	0.050301	0.949699	13970.17	0.986110972	6E-05	30042.34	5978441	99.64068	0.99005964
12	0.095522	0.904478	14633.67	0.986328262	6E-05	34040.11	6773982	112.8997	0.9915507
13	0.06035	0.93965	14635.27	0.986570527	6E-05	36556.97	7274847	121.2475	0.99055666
14	0.040252	0.959748	15115.74	0.986857748	6E-05	40630.01	8085373	134.7562	0.99055666
15	0.040252	0.959748	15257.14	0.986950158	6E-05	44476.87	8850903	147.5151	0.99105368
16	0.070399	0.929601	15574.31	0.986715386	6E-05	48814.78	9714154	161.9026	0.99105368
17	0.040252	0.959749	16061.62	0.987055057	6E-05	52596.1	10466632	174.4439	0.99055666
18	0.065374	0.934626	16315.92	0.986470624	6E-05	56323.53	11208402	186.8067	0.9915507
19	0.065374	0.934625	16508.11	0.986777826	6E-05	60294.17	11998554	199.9759	0.99105368
20	0.075424	0.924576	16817.97	0.986765338	6E-05	65534.99	13041471	217.3579	0.99105368
21	0.050301	0.949699	17031.53	0.986952656	6E-05	69008.7	13732733	228.8789	0.99055666
22	0.075423	0.924576	17151.65	0.986727874	6E-05	72677.69	14462863	241.0477	0.99105368
23	0.030203	0.969798	17133.17	0.987189926	6E-05	76695.43	15262409	254.3735	0.9915507
24	0.040252	0.959748	17515.32	0.987052559	6E-05	82230.77	16363928	272.7321	0.99055666
25	0.040252	0.959748	17703.49	0.987060052	6E-05	84932.23	16901520	281.692	0.99105368
26	0.040252	0.959748	17768.57	0.986850255	6E-05	88647.82	17640917	294.0153	0.99105368
27	0.040252	0.959748	17916.26	0.986650449	6E-05	93063.1	18519565	308.6594	0.9915507
28	0.025178	0.974822	18053.39	0.987227389	6E-05	97624.76	19427330	323.7888	0.99105368
29	0.020153	0.979847	18295.51	0.986795309	6E-05	101839.1	20265982	337.7664	0.99105368
30	0.045277	0.954723	18399.9	0.986585512	6E-05	106970.1	21287069	354.7845	0.99055666
31	0.055326	0.944674	18547.15	0.986750352	6E-05	112316.3	22350949	372.5158	0.99055666
32	0.035227	0.964773	18735.36	0.986910197	6E-05	117368.6	23356388	389.2731	0.99055666
33	0.025178	0.974821	18751.98	0.986767835	6E-05	119877.3	23855575	397.5929	0.99204771
34	0.025178	0.974822	18956.36	0.986905202	6E-05	123999.9	24675976	411.2663	0.99105368
35	0.035227	0.964773	18973.96	0.986960149	6E-05	128827.5	25636699	427.2783	0.99105368
36	0.065375	0.934625	19290.09	0.986745357	6E-05	133917.8	26649644	444.1607	0.99105368
37	0.015129	0.984871	19321.7	0.986940168	6E-05	138609.5	27583299	459.7217	0.99055666
38	0.030203	0.969797	19514.15	0.987334785	6E-05	143453.3	28547203	475.7867	0.99055666
39	0.025179	0.974821	19587.36	0.98708253	6E-05	147297.2	29312150	488.5358	0.99005964
40	0.035228	0.964772	19722.02	0.986790313	6E-05	152170.6	30281953	504.6992	0.9915507
41	0.055326	0.944674	20288.63	0.986960149	6E-05	160508.8	31941260	532.3543	0.99105368
42	0.025179	0.974822	20299.15	0.987542084	6E-05	166147.5	33063354	551.0559	0.9915507
43	0.030203	0.969797	20389.92	0.987399722	6E-05	169560	33742443	562.374	0.9915507
44	0.050302	0.949698	20286	0.987152462	6E-05	172557.2	34338875	572.3146	0.99105368
45	0.035228	0.964772	20769.8	0.98733978	6E-05	180348.7	35889392	598.1565	0.99055666
46	0.050302	0.949698	20710.6	0.987264853	6.00E-05	183306.5	28475139	474.5857	0.99055666
47	0.040253	0.959747	20967.33	0.987262355	6.00E-05	189763.4	37762919	629.382	0.99105368
48	0.025179	0.974821	21204.71	0.987564562	6.00E-05	198105.6	39423024	657.0504	0.9915507
49	0.0704	0.9296	21164.01	0.987239877	6.00E-05	200278	39855317	664.2553	0.99105368
50	0.035228	0.964772	21487.3	0.987437186	6.00E-05	206871.4	41167437	686.124	0.99055666
51	0.030204	0.969796	21826.14	0.987679451	6.00E-05	217693	43320919	722.0153	0.99055666
52	0.025179	0.974821	21922.72	0.987522104	6.00E-05	221511.2	44080724	734.6787	0.99105368
53	0.040253	0.959747	21895.37	0.987587041	6.00E-05	225604.6	44895313	748.2552	0.99105368
54	0.050302	0.949698	21945.7	0.987504621	6.00E-05	228949.1	45560879	759.348	0.99105368
55	0.030204	0.969796	22174.63	0.987866769	6.00E-05	247632.6	49278886	821.3148	0.99105368
56	0.050302	0.949698	22110.97	0.987562065	6.00E-05	252150.5	50177946	836.2991	0.9915507
57	0.075425	0.924575	22601.62	0.987492133	6.00E-05	262331.3	52203932	870.0655	0.99105368
58	0.0704	0.9296	22891.24	0.987806827	6.00E-05	270364.6	53802585	896.7098	0.99105368
59	0.065376	0.934624	22563.98	0.987651978	6.00E-05	271767.2	54081693	901.3616	0.99105368
60	0.055327	0.944673	22942.36	0.987651978	6.00E-05	280447.3	55809033	930.1506	0.99105368
61	0.065376	0.934624	23127.13	0.987656973	6.00E-05	289227.6	57556299	959.2716	0.99055666
62	0.055326	0.944673	22860.31	0.987612016	6.00E-05	289426.2	57595815	959.9302	0.9915507
63	0.085474	0.914526	23393.73	0.987839296	6.00E-05	299898.6	59679828	994.6638	0.99105368
64	0.060351	0.939648	23175.95	0.987607021	6.00E-05	301862.2	60070575	1001.176	0.99105368
65	0.105572	0.894428	23395.98	0.987569558	6.00E-05	308436.1	61378787	1022.98	0.99204771
66	0.105572	0.894428	23273.58	0.987384737	6.00E-05	311690.8	62026478	1033.775	0.99105368
67	0.090498	0.909502	23471.68	0.987509616	6.00E-05	319499.6	63580415	1059.674	0.99105368
68	0.085474	0.914526	23482.75	0.987836798	6.00E-05	324154.8	64506814	1075.114	0.99055666
69	0.130695	0.869305	23929.98	0.987374747	6.00E-05	333943.9	66454844	1107.581	0.99105368
70	0.140744	0.859256	23795.19	0.98740222	6.00E-05	339105.1	67481909	1124.698	0.99105368
71	0.075425	0.924575	23580.36	0.987477147	6.00E-05	341062.3	67871399	1131.19	0.99055666
72	0.145769	0.854231	23681.89	0.987442181	6.00E-05	347526.4	69157753	1152.629	0.9915507
73	0.196014	0.803986	23841.98	0.987119994	6.00E-05	353999.3	70445858	1174.098	0.99055666
74	0.185966	0.814035	24208.59	0.987244873	6.00E-05	362708.4	72178971	1202.983	0.99055666
75	0.170892	0.829108	23802.86	0.98724737	6.00E-05	363907.1	72417503	1206.958	0.99055666
76	0.160842	0.839158	24353.72	0.987334785	6.00E-05	375980.6	74820141	1247.002	0.99055666
77	0.185966	0.814035	24083.91	0.986957651	6.00E-05	374982.4	74621492	1243.692	0.99105368
78	0.206063	0.793937	24215.16	0.987192423	6.00E-05	381791.1	75976421	1266.274	0.99055666
79	0.160842	0.839158	24249.91	0.987127486	6.00E-05	393548.7	78316197	1305.27	0.99105368
80	0.155818	0.844182	24367.54	0.987075037	6.00E-05	393758.5	78357947	1305.966	0.99055666

这次实验共持续了超过4.5万分钟，共收集了81组数据。卷积核数量=0就是三层的无核的网络。将pave曲线画成图

 

这是卷积核从0到80的图，可以观察到这个图有一个顶点的。顶点为55。

 

这是卷积核从10到55的图，网络的性能随着卷积核的数量的增加而增加。7*7卷积核的上升区间有55个核，对应5*5卷积核的上升区间为16个核，3*3卷积核的上升区间为4个核。

这个现象清楚的表明在网络结构不变的前提下，卷积核越大对应的上升区间越大。

 

这是卷积核从40到80的图，这张图的第三高的峰出现在68个核，68与55两者的对应的性能数值差异仅为0.000029。表明网络性能下降有一个大约为13个核的缓冲区间。但当卷积核的数量大于68以后网络的性能迅速下降。

因此到底应该用3*3的卷积核还是用5*5的卷积核还是7*7的卷积核？

	3*3	5*5	7*7	5*5/3*3	7*7/5*5
	平均准确率p-ave	平均准确率p-ave	平均准确率p-ave
	6.00E-05	6.00E-05	6.00E-05
0	0.981171	0.981171	0.981171	1	1
1	0.975916	0.978588	0.976048	1.002738	0.997405
2	0.981326	0.983376	0.981191	1.002089	0.997778
3	0.983633	0.985159	0.983136	1.001551	0.997947
4	0.983651	0.986268	0.98426	1.00266	0.997964
5	0.983289	0.986143	0.984795	1.002903	0.998633
6	0.983506	0.986323	0.986064	1.002864	0.999737
7	0.982744	0.986605	0.986086	1.003929	0.999474
8	0.982694	0.98689	0.986218	1.00427	0.999319
9	0.981885	0.98697	0.986083	1.005179	0.999102
10	0.980983	0.986988	0.985991	1.006121	0.99899
11	0.981401	0.987225	0.986111	1.005934	0.998872
12	0.98214	0.986988	0.986328	1.004936	0.999332
13		0.987295	0.986571		0.999266
14		0.987132	0.986858		0.999722
15		0.987065	0.98695		0.999884
16		0.987322	0.986715		0.999386
17		0.987227	0.987055		0.999826
18		0.98672	0.986471		0.999747
19		0.987137	0.986778		0.999636
20		0.986988	0.986765		0.999774
21		0.986855	0.986953		1.000099
22		0.986241	0.986728		1.000494
23		0.986468	0.98719		1.000732

 

7*7卷积核的上升区间为55，5*5卷积核的上升区间为16。当卷积核数量大于20以后7*7卷积核的性能优于5*5卷积核。

2分类	3*3	5*5	7*7
性能上升区间	4	16	55
max-p-ave	0.9838731	0.987322	0.987867
耗时min/199次	11.6074	102.3604	821.3148

 

至少仅就这三次实验来说，卷积核越大网络的性能上升区间越大，上升区间越大性能峰值越大。但代价是卷积核越多网络越慢，7*7卷积核的性能峰值比5*5卷积核峰值要好0.5‰，但为了这万分之5要多花8倍的时间。