卷积核的数量是不是越多越好?-分类0,5

 

制作一个二分类的网络来分类mnist的0和5,并向网络上加卷积核从1个核到9个核。

网络的结构是

(mnist 0 ,mnist9)81-con(3*3)*n-(49*n)-30-2-(1,0) || (0,1)

将mnist的28*28的图片压缩到9*9,用n个3*3的卷积核,节点数分别为n*49,30,2。让0向(1,0)收敛,让5向(0,1)收敛,让n分别等于1-9.

与Tensorflow不同的是这个网络通过网络的输出值与目标函数的绝对误差判断是否停止:

if (Math.abs(网路输出值[0]-目标函数[0])< δ  &&  Math.abs(网络输出值[1]-目标函数[1])< δ   ),

其中δ分别等于0.5到1e-6.的34个值。对应每个δ收敛199次,分别记录与之对应的迭代次数,收敛时间,并计算199次的平均准确率和199次的最大准确率,来比较这10个网络的性能差异。

 

首先比较最大准确率

 

 

81-30-2

1

2

3

4

5

6

7

8

9

δ

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

0.5

0.7948718

0.6730769

0.7692308

0.7638889

0.8178419

0.7820513

0.8012821

0.9027778

0.8904915

0.767094

0.4

0.9428419

0.9636752

0.9604701

0.9620727

0.9529915

0.9583333

0.9610043

0.9636752

0.9465812

0.9519231

0.3

0.9524573

0.9711538

0.9679487

0.9594017

0.9599359

0.9513889

0.9487179

0.9567308

0.9444444

0.9428419

0.2

0.9626068

0.9647436

0.9604701

0.9604701

0.9535256

0.9465812

0.9481838

0.9540598

0.9508547

0.9444444

0.1

0.9674145

0.9674145

0.9647436

0.9652778

0.957265

0.957265

0.9540598

0.9503205

0.954594

0.9449786

0.01

0.9695513

0.9700855

0.971688

0.9684829

0.9679487

0.9668803

0.9684829

0.9732906

0.9706197

0.971688

0.001

0.9775641

0.9727564

0.9797009

0.9807692

0.9807692

0.9807692

0.9786325

0.9818376

0.9823718

0.980235

9.00E-04

0.9775641

0.9738248

0.9813034

0.9818376

0.9807692

0.9813034

0.9807692

0.9823718

0.9818376

0.9818376

8.00E-04

0.9775641

0.9754274

0.9807692

0.9813034

0.9823718

0.9823718

0.982906

0.9823718

0.982906

0.9818376

7.00E-04

0.9770299

0.971688

0.9818376

0.982906

0.9834402

0.982906

0.982906

0.9818376

0.982906

0.9834402

6.00E-04

0.9770299

0.9727564

0.982906

0.982906

0.9839744

0.9834402

0.9823718

0.9839744

0.9834402

0.9823718

5.00E-04

0.9759615

0.9754274

0.9823718

0.9839744

0.9850427

0.9823718

0.9839744

0.9834402

0.9839744

0.9839744

4.00E-04

0.974359

0.9748932

0.9834402

0.982906

0.9839744

0.9839744

0.9845085

0.9845085

0.9850427

0.9850427

3.00E-04

0.980235

0.9770299

0.9845085

0.9866453

0.9845085

0.9839744

0.9855769

0.9850427

0.9855769

0.9850427

2.00E-04

0.980235

0.9754274

0.9855769

0.9861111

0.9866453

0.9850427

0.9850427

0.9861111

0.9861111

0.9866453

1.00E-04

0.9807692

0.9786325

0.9866453

0.9877137

0.9866453

0.9850427

0.9861111

0.9866453

0.9861111

0.9850427

9.00E-05

0.9807692

0.9775641

0.9861111

0.9871795

0.9877137

0.9871795

0.9855769

0.9866453

0.9861111

0.9855769

8.00E-05

0.9807692

0.9764957

0.9887821

0.9882479

0.9871795

0.9882479

0.9866453

0.9877137

0.9861111

0.9855769

7.00E-05

0.9807692

0.9775641

0.9887821

0.9877137

0.9877137

0.9903846

0.9871795

0.9877137

0.9871795

0.9871795

6.00E-05

0.9813034

0.9791667

0.9882479

0.9877137

0.9887821

0.9903846

0.9882479

0.9877137

0.9866453

0.9871795

5.00E-05

0.9807692

0.9780983

0.9898504

0.9887821

0.9882479

0.9898504

0.9882479

0.9877137

0.9866453

0.9882479

4.00E-05

0.9797009

0.9775641

0.9893162

0.9898504

0.9898504

0.9903846

0.9887821

0.9887821

0.9866453

0.9866453

3.00E-05

0.9797009

0.980235

0.9898504

0.9887821

0.9893162

0.9925214

0.9887821

0.9877137

0.9882479

0.9882479

2.00E-05

0.9791667

0.9813034

0.991453

0.9887821

0.9909188

0.991453

0.9898504

0.9882479

0.9882479

0.9877137

1.00E-05

0.9813034

0.9845085

0.9903846

0.9903846

0.9903846

0.991453

0.9909188

0.9887821

0.9882479

0.9893162

9.00E-06

0.9813034

0.9845085

0.9909188

0.991453

0.9903846

0.9919872

0.9893162

0.9893162

0.9893162

0.9903846

8.00E-06

0.9813034

0.982906

0.991453

0.991453

0.9903846

0.9925214

0.9893162

0.9893162

0.9893162

0.991453

7.00E-06

0.9813034

0.9850427

0.9909188

0.9925214

0.9909188

0.9919872

0.991453

0.9898504

0.9898504

0.9909188

6.00E-06

0.9818376

0.9839744

0.9893162

0.991453

0.9925214

0.991453

0.9893162

0.9893162

0.9909188

0.9909188

5.00E-06

0.9818376

0.9850427

0.991453

0.9935897

0.991453

0.9925214

0.9909188

0.9903846

0.9903846

0.9893162

4.00E-06

0.9813034

0.9866453

0.9935897

0.9930556

0.9930556

0.9925214

0.9919872

0.9898504

0.991453

0.9898504

3.00E-06

0.9834402

0.9855769

0.9930556

0.9930556

0.9919872

0.9919872

0.9925214

0.9909188

0.9903846

0.9909188

2.00E-06

0.9861111

0.9871795

0.9951923

0.9935897

0.9925214

0.9930556

0.991453

0.9909188

0.9925214

0.9919872

1.00E-06

0.9882479

0.9893162

0.9951923

0.9946581

0.9930556

0.9935897

0.9941239

0.9903846

0.9909188

 

 

卷积核的数量是不是越多越好?-分类0,5_第1张图片

卷积核的数量是不是越多越好?-分类0,5_第2张图片

 

2>3>4>5>6>7>8>9>1>81-30-2

随着卷积核数量的增加网络的最大性能先上升,当n=2时最大性能到达顶点,然后随着卷积核数量的增加最大性能开始下降。最大准确率曲线是一条开口向下有极大值的曲线。原始的未加卷积核的网络81-30-2的最大性能小于1个卷积核的网络。

 

2.比较平均性能

 

81-30-2

1

81-30-2

2

3

4

5

6

7

8

9

δ

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

0.5

0.5335518

0.5023891

0.5335518

0.5185033

0.5139963

0.5152579

0.5164525

0.5226802

0.5270476

0.5192549

0.5158699

0.4

0.9188652

0.715718

0.9188652

0.7907755

0.7566518

0.7538386

0.7368949

0.7321302

0.7023123

0.7222598

0.6953357

0.3

0.9282497

0.8465807

0.9282497

0.8974064

0.882833

0.8699158

0.8699427

0.8521937

0.8459364

0.8328636

0.8368579

0.2

0.9465731

0.8818961

0.9465731

0.9070137

0.9074754

0.9003296

0.8978654

0.891082

0.8898338

0.8887493

0.8770884

0.1

0.9638148

0.904426

0.9638148

0.925815

0.9159687

0.9124039

0.9085143

0.9077116

0.9094189

0.9066137

0.8981714

0.01

0.9656402

0.9115181

0.9656402

0.9400233

0.9378436

0.9272887

0.9153996

0.9101195

0.8945395

0.8760925

0.8632318

0.001

0.9733148

0.903261

0.9733148

0.9617666

0.9538988

0.9504923

0.9438888

0.9492763

0.9436687

0.9433546

0.9410273

9.00E-04

0.9732342

0.9092202

0.9732342

0.9617237

0.959391

0.9572381

0.9567335

0.9579441

0.9562127

0.9571683

0.957469

8.00E-04

0.9731752

0.9105597

0.9731752

0.9651006

0.9609909

0.9641101

0.9621747

0.9644993

0.9630471

0.963235

0.9646228

7.00E-04

0.9728396

0.9110295

0.9728396

0.9668911

0.9688614

0.9664025

0.9659999

0.9659167

0.9660213

0.9664857

0.9644107

6.00E-04

0.972088

0.9108899

0.972088

0.9669609

0.9673796

0.9678467

0.9679487

0.9702761

0.968491

0.9676266

0.9660562

5.00E-04

0.9712746

0.91475

0.9712746

0.9700989

0.9727215

0.9699244

0.9717149

0.9708988

0.9723618

0.9699405

0.9666226

4.00E-04

0.9715672

0.9135072

0.9715672

0.9737657

0.9747079

0.9734731

0.9730758

0.971782

0.9713095

0.969111

0.9680266

3.00E-04

0.9745925

0.9094457

0.9745925

0.9743348

0.9740825

0.9747589

0.9730946

0.973347

0.9713632

0.9704801

0.9671756

2.00E-04

0.9764313

0.9190611

0.9764313

0.9768259

0.9767105

0.9756475

0.9741845

0.97361

0.9727269

0.9698841

0.9705338

1.00E-04

0.974861

0.9286497

0.974861

0.9799156

0.9793089

0.9780527

0.9642336

0.9778218

0.9756877

0.9740288

0.9741818

9.00E-05

0.9772608

0.9331299

0.9772608

0.9800069

0.9808605

0.9800767

0.9738677

0.9778057

0.9762783

0.9761038

0.9755294

8.00E-05

0.9783614

0.9257693

0.9783614

0.9808605

0.9806753

0.9799559

0.9765226

0.9771158

0.9773332

0.9777574

0.9743321

7.00E-05

0.9792284

0.9319353

0.9792284

0.9824067

0.9817651

0.9803397

0.9770541

0.9789546

0.9781359

0.9778862

0.9747938

6.00E-05

0.9795801

0.9219656

0.9795801

0.9829301

0.982557

0.9810135

0.9779963

0.9785681

0.9776339

0.9779855

0.9760931

5.00E-05

0.9791157

0.9300589

0.9791157

0.9827771

0.9829543

0.9817973

0.9796579

0.979172

0.9777976

0.978509

0.9770031

4.00E-05

0.9788687

0.9251224

0.9788687

0.9837918

0.9830295

0.9820577

0.9800901

0.9799666

0.9777332

0.9780849

0.9777252

3.00E-05

0.9788338

0.9216435

0.9788338

0.9843529

0.9833382

0.982761

0.9820014

0.9799344

0.9789143

0.9786701

0.978458

2.00E-05

0.9787935

0.9209053

0.9787935

0.9851286

0.9840307

0.9831154

0.9828738

0.9804498

0.9805196

0.9794861

0.9773708

1.00E-05

0.9783506

0.9308052

0.9783506

0.9853568

0.9842133

0.983816

0.9837516

0.9821114

0.980788

0.9797116

0.9699244

9.00E-06

0.9782352

0.9433573

0.9782352

0.9860467

0.9852763

0.9841811

0.9850052

0.982251

0.9801438

0.9727725

0.9775775

8.00E-06

0.9781332

0.9331621

0.9781332

0.9865191

0.9859098

0.9846642

0.9850266

0.9828496

0.9793895

0.979062

0.9805652

7.00E-06

0.9782218

0.9447344

0.9782218

0.9865809

0.9862749

0.9853568

0.9853326

0.9829677

0.9815289

0.9789224

0.9794861

6.00E-06

0.9785761

0.9488844

0.9785761

0.977454

0.9866238

0.985687

0.985789

0.9829946

0.981145

0.979419

0.9815343

5.00E-06

0.9790861

0.9555389

0.9790861

0.9850374

0.9871016

0.9865594

0.9858937

0.9835368

0.9823423

0.98078

0.9820201

4.00E-06

0.9797331

0.9611895

0.9797331

0.9869164

0.9868413

0.9859554

0.9862158

0.9849837

0.9821651

0.9823557

0.9758005

3.00E-06

0.9802726

0.9644107

0.9802726

0.9879257

0.9874587

0.9862212

0.9863312

0.9837865

0.9817812

0.9826214

0.980525

2.00E-06

0.9802673

0.9657878

0.9802673

0.9891149

0.9884841

0.9874587

0.9860682

0.9850749

0.9843126

0.9839502

0.9819557

1.00E-06

0.9832684

0.9736745

0.9832684

0.9895203

0.9887123

0.9878694

0.9870292

0.9842428

0.9839717

0.9848495

 

卷积核的数量是不是越多越好?-分类0,5_第3张图片

卷积核的数量是不是越多越好?-分类0,5_第4张图片

 

2>3>4>5>6>7>8>9>81-30-2>1

平均性能也是一条开口向下有极大值的曲线,顶点在n=2,当n>2后随着卷积核的数量的增加网络的平均性能下降,原始的无卷积核的网络81-30-2的平均性能在9与1之间。

 

3.比较迭代次数

 

81-30-2

1

2

3

4

5

6

7

8

9

δ

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

0.5

9.160804

18.492462

17.311558

15.743719

15.432161

16.346734

16.798995

15.236181

12.527638

13.844221

0.4

259.27136

2288.4774

298.56281

236.08543

220.34171

208.94472

193.33668

194.93467

195.60302

173.17588

0.3

344.46734

2537.2513

154.73367

144.78392

140.98995

135.54774

139.46734

136.20603

131.78392

131.72864

0.2

447.94472

2826.4472

163.06533

145.31156

138.14573

142.52261

136.78392

142.35678

140.75377

138.37688

0.1

546.02513

2872.4774

182.70352

173.42714

163.88442

170.22111

166.94975

168.49749

167.97487

173.1005

0.01

855.83417

3587.7035

444.98995

492.43719

505.8794

518.46734

518.61307

526.22111

532.85427

519.0201

0.001

1523.0955

4529.1508

1436.4573

1692.8995

1901.6985

1920.9749

1954.9246

1907.5427

1998.1508

2039.1859

9.00E-04

1575.8392

4472.7688

1420.0452

1642.4271

1883.0754

1885.6432

1929.2211

1934.4422

2052.6181

2154.0955

8.00E-04

1657.0251

4529.0201

1491.2211

1749.6633

1986.2864

1940.6734

2014.0603

1949.593

2134.1055

2116.809

7.00E-04

1715.8492

4662.1859

1720.3116

1978.9397

2035.6985

2038.5678

2104

2170.8995

2200.1709

2291.8492

6.00E-04

1784.804

4733.206

1785.5176

2072.6784

2165.9045

2188.9648

2285.201

2195.0503

2285.4774

2328.201

5.00E-04

1845.3668

4991.0704

1883.0553

2268.2965

2208.1809

2327.2211

2576.4874

2318.2412

2525.1106

2651.4774

4.00E-04

1991.608

5026.5327

2050.1407

2512.005

2483.1608

2693.8141

2735.6633

2599.0804

2686.9799

2738.3216

3.00E-04

2476.593

5324.7889

2454.9648

2812.5779

2978.3568

2997.9598

3013.9196

3046.5678

2939.8643

3272.7186

2.00E-04

2879.5678

5625.8291

2968.9749

3547.2362

3372.3266

3704.4573

3738.206

3535.6583

3900.3668

3865.3216

1.00E-04

3656.4422

6300.1558

4395.1256

4725.9799

4806.7839

5709.2111

5251.6784

5196.9196

5116.0603

5670.2764

9.00E-05

3871.3367

6393.1809

4249.2563

4417.4121

4810.5578

5519.6834

5046.7487

5074.6633

5158.3317

5383.608

8.00E-05

4092.3317

6461.5829

4191.804

4866.4171

4839.5729

5669.0302

5343.5427

5026.0251

5428.9749

5520.8291

7.00E-05

4245.1759

6737.6583

4436.8945

4731.3266

5093.0653

5754.7688

5417.1307

5070.9347

5305.6784

5715.3668

6.00E-05

4380.0302

7093.3367

4591.5427

4900.6734

4905.6784

6238.9397

5270.2764

5661.1055

5813.1759

5918.1759

5.00E-05

4537.1457

7531.8392

4661.5176

5124.5126

5510.2161

6595.4372

5685.5729

5880.8844

6158.4221

6208.3467

4.00E-05

4640.3417

7936.4372

4891.402

5516.794

6135.8945

7171.2362

6050.5879

6745.5779

6235.9447

6859.3417

3.00E-05

4661.799

9152.593

5546.593

6137.5879

6695.1457

8369.5477

6584.7889

6960.005

7222.0101

7973.3216

2.00E-05

4680.2814

11076.392

6096.0151

7011.6583

7930.3819

10117.513

7742.3618

8165.598

9197.5678

9313.9598

1.00E-05

5416.9347

21376.387

7696.7487

9081.0251

10953.553

13592.714

10608.905

11073.864

12293.543

26256.166

9.00E-06

5648.5528

23626.186

7388.9196

9121.7035

11644.558

13399.03

10421.492

11290.291

25672.563

28369.141

8.00E-06

5854.5126

22841.116

7448.0352

10441.879

12020.628

14497.573

10589.693

12226.497

24378.653

31408.497

7.00E-06

6466.8241

27217.905

8028.1106

10264.025

11424.136

15736

11804.558

12377.161

28071.392

32087.392

6.00E-06

7063.5276

29178.518

13824.156

10173.643

12465.075

14446.613

12330.91

13247.849

27945.513

31361.045

5.00E-06

8079.7688

32877.06

17375.497

10970.327

13638.372

15564.558

13761.432

15111.543

28023.92

31892.99

4.00E-06

9789.3668

38820.513

18059.628

10907.543

12893.543

16265.276

13874.136

15322.347

30350.889

43566.864

3.00E-06

12842.985

42096.995

17814.714

12538.608

14224.08

18842.116

15662.533

17281.513

33741

44026.899

2.00E-06

15130.764

45979.095

19888.558

14292.839

18005.352

23446.09

19432.312

20531.422

37886.804

53454.829

1.00E-06

22746.784

57660.256

22674.513

18676.844

25511.854

31373.98

46643.784

25671.94

58280.683

 

 

卷积核的数量是不是越多越好?-分类0,5_第5张图片

1>2>81-30-2>3<4<5<6<8<9

除了第7条曲线迭代次数大体上是开口向上有极小值的曲线,当n=3时有极小值,原始网络的迭代次数在2和3之间。

 

4.最后比较收敛时间

 

81-30-2

1

2

3

4

5

6

7

8

9

δ

耗时 min/199

耗时 min/200

耗时 min/201

耗时 min/202

耗时 min/203

耗时 min/204

耗时 min/205

耗时 min/206

耗时 min/207

耗时 min/208

0.5

3.74165

2.81195

2.3356167

2.3812833

2.41515

2.9003

2.6563167

2.6037

2.84575

2.8494333

0.4

4.1931333

5.6846

2.7243167

2.70495

2.7242

3.2568667

2.9844

2.9688167

3.2323833

3.3662333

0.3

4.3721667

6.0041333

2.5128

2.6714667

2.6016333

3.1050333

2.8730167

2.8513333

3.10375

3.19675

0.2

1.8564

6.3734667

2.5300833

2.5535333

2.58635

3.0920333

2.86965

2.8619667

3.1330667

1.5324833

0.1

4.6676167

6.4437667

2.5244667

2.58445

2.62775

3.1427833

2.9340667

2.9192333

3.1962167

3.3199833

0.01

5.2088833

7.3604833

2.8403333

-0.58325

3.2821333

3.7929667

3.6354833

3.6760833

4.0656333

4.34505

0.001

6.3676667

8.5501167

4.1040667

4.81455

5.6036

6.4232333

6.4751

6.5676833

4.7247667

8.4573333

9.00E-04

6.4966833

8.4807667

4.0384333

4.78085

4.4528667

6.3847167

6.43075

6.62765

7.4476667

8.8073

8.00E-04

6.7034

8.5498333

4.0271167

4.902

5.7288833

6.3719

6.6306333

6.7841833

7.6419833

8.7222167

7.00E-04

6.75745

8.7181

4.40325

5.2740167

5.7882833

6.6370667

4.7293

7.3665333

7.8004667

9.1840167

6.00E-04

6.8404

8.80775

4.4865833

5.37725

5.97655

6.9272667

7.19365

7.2842667

7.9979

9.2548

5.00E-04

6.96285

9.1413333

4.6114333

5.5824167

6.0483333

4.6591833

7.75675

7.63485

8.5834667

4.68755

4.00E-04

7.2862

9.1793667

2.0086833

5.9000167

6.5423833

7.85475

8.1671333

8.3504167

8.9637333

10.353083

3.00E-04

5.5816667

9.5621333

5.3298333

6.3522167

7.35655

8.4423167

8.7433667

9.4013667

9.5543167

11.735717

2.00E-04

8.71945

9.94355

5.9963833

7.6161167

7.98195

9.8058333

10.35895

7.4389833

11.78

12.985167

1.00E-04

10.0387

10.8083

7.8396667

6.6227

8.08105

13.075883

10.481833

13.735567

14.89045

18.13725

9.00E-05

10.4178

10.913867

7.6425667

8.9246333

10.598233

12.833733

13.65385

13.224017

15.28805

12.19625

8.00E-05

10.768333

11.00325

7.6148167

9.4912

10.727383

13.143783

14.537217

13.125967

12.697583

17.459533

7.00E-05

11.03965

11.351883

7.8541833

9.2809

11.121167

13.43135

14.232983

13.212067

15.641267

17.93935

6.00E-05

10.08165

11.803567

8.0594

9.40985

10.833517

11.598917

13.148567

14.464867

16.658133

14.880983

5.00E-05

11.744183

11.558367

8.1961

9.7128333

11.570617

15.025217

9.5606

14.932633

17.066967

19.808217

4.00E-05

11.89545

13.316783

6.1954

10.30215

9.7710167

16.06235

14.383

16.738283

15.5816

21.337733

3.00E-05

11.962133

15.221267

9.3110833

8.50165

13.173067

15.182917

15.406933

15.875433

19.50825

17.410333

2.00E-05

11.947567

17.901783

10.009533

12.496633

15.301883

21.597833

17.651783

19.716983

24.296367

25.427117

1.00E-05

11.3564

28.662117

12.06555

15.452733

20.2102

27.94395

23.37535

25.817217

30.363617

66.972917

9.00E-06

13.647017

32.964433

11.656317

15.563267

19.6612

24.301083

23.03795

26.32625

62.753017

73.75985

8.00E-06

14.8104

31.809883

9.7350667

14.798517

21.876733

29.526217

20.793

29.18205

56.518383

78.87625

7.00E-06

19.61805

36.461317

12.50245

16.808383

19.1992

28.709183

25.774433

27.901667

65.33635

82.823083

6.00E-06

20.565867

32.756467

19.5214

16.716867

22.1395

29.47665

25.1147

32.154583

66.056617

78.91485

5.00E-06

20.4579

39.824267

23.790033

17.896783

23.918717

28.025917

30.541083

33.241817

67.500017

81.8001

4.00E-06

26.225617

43.384617

25.791183

17.6785

24.153483

31.654933

28.107583

34.056233

71.296233

112.10312

3.00E-06

30.387583

47.09945

25.603317

19.945967

26.7583

35.474967

34.3632

39.994767

80.6607

113.52042

2.00E-06

38.4815

50.440683

23.894317

20.51575

30.430383

43.04565

39.376083

46.21565

87.64325

148.69013

1.00E-06

52.468967

64.474033

32.313933

28.640433

46.072367

57.658567

93.034733

56.501933

133.82317

 

 

卷积核的数量是不是越多越好?-分类0,5_第6张图片

1>81-30-2>2>3<4<5<6<8<9

与迭代次数类似曲线也是开口向上的曲线,当n=3时耗时最少

 

所以综合上面的4个表格

  1. 增加卷积核网络最大性能先升后降有极大值,超过极大值随着卷积核数量的增加最大性能下降
  2. 增加卷积核网络的平均性能也先升后降有极大值。超过极大值网络的平均性能同样会下降
  3. 迭代次数曲线近似为一条开口向上的曲线有极小值,这个极小值对应的n与1和2的极大值对应的n接近
  4. 收敛时间与迭代次数有正比关系

 

最大性能:2>3>4>5>6>7>8>9>1>81-30-2

平均性能:2>3>4>5>6>7>8>9>81-30-2>1

迭代次数:1>2>81-30-2>3<4<5<6<8<9

收敛时间:1>81-30-2>2>3<4<5<6<8<9

因此对于这个网络无论更在乎最大性能还是平均性能都应该选择2个卷积核,因为n=2同时是这个网络在1-9个卷积核内平均性能和最大性能的极大值。

如果更在乎收敛效率也可以选择3个卷积核,3比2的性能稍差但需要的迭代次数只有2的82%,可以节省些时间。

如果选择了8个卷积核当收敛标准δ=1e-6的时候要比2个卷积核多付出2.57倍的计算量但最大性能比n=2的网络还是要差0.5%.因此就这个网络来说卷积核的数量肯定不是越多越好,当卷积核的数量超过2个以后卷积核的数量对网络的性能已经没有任何正面价值,而且数量越多越慢性能也越差。

 

实验数据

学习率r=0.1

权重初始化方式

Random rand1 =new Random();

int ti1=rand1.nextInt(98)+1;

int xx=1;

if(ti1%2==0)

{ xx=-1;}

tw[a][b]=xx*((double)ti1/x);

第一层第二层和卷积核的权重的初始化的x分别为1000,1000,200

 

 

 

你可能感兴趣的:(用神经网络模拟分子)