What are the possible ways to increase the accuracy while training through CNN network?

Question

Ravish Raj 2017-6-16

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/345029-what-are-the-possible-ways-to-increase-the-accuracy-while-training-through-cnn-network

I am trying to pass 4-D image arrays through the CNN layers , i.e. through this : layers = [imageInputLayer([32 32 1]) convolution2dLayer(5,20) reluLayer() maxPooling2dLayer(2,'Stride',2) fullyConnectedLayer(size(categories(trainAngle),1)) softmaxLayer classificationLayer]; and I have images,already gray scaled. and I am also passing, the corresponding angle values which has the dimension of [size(Images,4) 1], i.e. 2-D angle array! following lines of codes are : options = trainingOptions('sgdm', 'MaxEpochs', 100, ... 'InitialLearnRate', 0.001);

convnet = trainNetwork(trainZ, trainAngle, layers,options);

where trainZ is training images 4-D & trainAngle is training angle of 2-D size.

 resultant_Train = classify(convnet,trainZ); %Training data
 resultant_Valid = classify(convnet,validZ);
where validZ is validation or testing data.

But my accuracy chart was :- ========================================================================================= | Epoch | Iteration | Time Elapsed | Mini-batch | Mini-batch | Base Learning| | | | (seconds) | Loss | Accuracy | Rate =========================================================================================| | 1 | 1 | 0.34 | 7.0118 | 0.00% | 0.0010 2 | 50 | 16.21 | 3.4001 | 25.78% | 0.0010 | | 4 | 100 | 32.37 | 2.5385 | 39.06% | 0.0010 | | 5 | 150 | 48.33 | 2.8044 | 41.41% | 0.0010 | | 7 | 200 | 64.35 | 2.8276 | 40.63% | 0.0010 | | 9 | 250 | 80.58 | 2.7260 | 39.84% | 0.0010 | | 10 | 300 | 96.86 | 2.5052 | 49.22% | 0.0010 | | 12 | 350 | 113.49 | 3.0974 | 45.31% | 0.0010 | | 14 | 400 | 129.63 | 2.8960 | 42.19% | 0.0010 | | 15 | 450 | 145.80 | 2.2082 | 53.13% | 0.0010 | | 17 | 500 | 161.91 | 1.6596 | 67.19% | 0.0010 | | 19 | 550 | 178.04 | 2.5425 | 57.03% | 0.0010 | | 20 | 600 | 194.35 | 4.2116 | 49.22% | 0.0010 | | 22 | 650 | 210.80 | 2.2605 | 56.25% | 0.0010 | | 24 | 700 | 227.26 | 2.3513 | 60.94% | 0.0010 | | 25 | 750 | 243.27 | 2.8299 | 60.16% | 0.0010 | | 27 | 800 | 259.24 | 2.4145 | 64.06% | 0.0010 | | 29 | 850 | 275.41 | 1.6220 | 67.19% | 0.0010 | | 30 | 900 | 291.33 | 2.6989 | 61.72% | 0.0010 | | 32 | 950 | 307.32 | 2.5999 | 67.19% | 0.0010 | | 34 | 1000 | 323.28 | 3.7991 | 56.25% | 0.0010 | | 35 | 1050 | 339.20 | 2.1487 | 70.31% | 0.0010 | | 37 | 1100 | 355.05 | 2.7791 | 68.75% | 0.0010 | | 39 | 1150 | 370.92 | 4.8829 | 53.13% | 0.0010 | | 40 | 1200 | 386.86 | 2.0481 | 69.53% | 0.0010 | | 42 | 1250 | 402.82 | 2.5822 | 71.88% | 0.0010 | | 44 | 1300 | 418.93 | 2.9441 | 65.63% | 0.0010 | | 45 | 1350 | 434.86 | 2.9613 | 64.06% | 0.0010 | | 47 | 1400 | 450.90 | 2.5159 | 71.88% | 0.0010 | | 49 | 1450 | 466.87 | 1.2781 | 77.34% | 0.0010 | | 50 | 1500 | 482.75 | 2.0612 | 68.75% | 0.0010 | | 52 | 1550 | 498.81 | 1.7994 | 81.25% | 0.0010 | | 54 | 1600 | 515.03 | 2.8094 | 73.44% | 0.0010 | | 55 | 1650 | 531.58 | 2.2264 | 73.44% | 0.0010 | | 57 | 1700 | 548.05 | 2.2138 | 77.34% | 0.0010 | | 59 | 1750 | 564.88 | 1.9252 | 74.22% | 0.0010 | | 60 | 1800 | 581.93 | 1.6136 | 82.81% | 0.0010 | | 62 | 1850 | 598.06 | 2.9149 | 69.53% | 0.0010 | | 64 | 1900 | 614.52 | 1.0750 | 82.03% | 0.0010 | | 65 | 1950 | 630.70 | 2.3359 | 78.91% | 0.0010 | | 67 | 2000 | 646.65 | 1.4321 | 83.59% | 0.0010 | | 69 | 2050 | 662.65 | 2.2901 | 75.78% | 0.0010 | | 70 | 2100 | 678.92 | 2.5174 | 78.13% | 0.0010 | | 72 | 2150 | 695.66 | 2.2956 | 78.13% | 0.0010 | | 74 | 2200 | 712.21 | 2.0279 | 81.25% | 0.0010 | | 75 | 2250 | 729.02 | 3.7779 | 68.75% | 0.0010 | | 77 | 2300 | 746.28 | 1.8612 | 82.03% | 0.0010 | | 79 | 2350 | 762.83 | 4.6060 | 61.72% | 0.0010 | | 80 | 2400 | 779.21 | 1.6944 | 85.16% | 0.0010 | | 82 | 2450 | 796.09 | 3.1484 | 75.78% | 0.0010 | | 84 | 2500 | 812.81 | 3.3044 | 72.66% | 0.0010 | | 85 | 2550 | 829.76 | 2.3785 | 81.25% | 0.0010 | | 87 | 2600 | 846.00 | 4.7383 | 63.28% | 0.0010 | | 89 | 2650 | 861.98 | 3.1656 | 71.88% | 0.0010 | | 90 | 2700 | 877.81 | 2.6156 | 75.78% | 0.0010 | | 92 | 2750 | 893.65 | 2.3809 | 80.47% | 0.0010 | | 94 | 2800 | 909.45 | 3.8142 | 68.75% | 0.0010 | | 95 | 2850 | 925.28 | 2.7810 | 77.34% | 0.0010 | | 97 | 2900 | 941.25 | 3.7744 | 73.44% | 0.0010 | | 99 | 2950 | 957.26 | 2.8002 | 78.91% | 0.0010 | | 100 | 3000 | 973.10 | 4.2253 | 70.31% | 0.0010 | ========================================================================================= And my training accuracy is 66% and validation accuracy is 6% only. How can I increase my accuracy, Because it is too bad. I have split the data in 75:25 for training and testing purpose.