[人工智能]Compared the network layer

Problems.

We found shallow CNN worked better in our target detection ,however, one guy told me "your result is not consistant with the traditional result."

Therefore, we retrain our network.

We did found that all the loss values of cnn model are?convergence.

?Compared layers

Test layer
analyzeNetwork(detector_Faster_res50.Network)

%resnet 50
analyzeNetwork( net50)

analyzeNetwork(detector_yolo_res50.Network)

Res50?

faster RCNN+Res50

layersToRemove = {
    'fc1000'
    'fc1000_softmax'
    'ClassificationLayer_fc1000'
    };

lgraph = removeLayers(lgraph1, layersToRemove);

yolo res50

?

Resnet18

?

?

?

?

?yolo-res18

analyzeNetwork(detector_yolo_res18.Network)

inputSize = [224 224 3];
numClasses = 1;
featureExtractionNetwork = resnet18;
featureLayer = 'res5b_relu';
lgraphYoloRes18 = yolov2Layers(inputSize,numClasses,anchorBoxes,featureExtractionNetwork,featureLayer);
figure,plot(lgraphYoloRes18)
analyzeNetwork(lgraphYoloRes18)

?

?

?Fast Res18

%build fast resnet18
lgraphFastRes18 = fasterRCNNLayers(inputSize,numClasses,anchorBoxes,featureExtractionNetwork,featureLayer);
analyzeNetwork(lgraphFastRes18)

?

3 Retrain

Retrain all lthe model to let us know whether our previous training is not good.?

because ,we found shallow cnn is better.

3.1 fast+res50

inputSize = [224 224 3];

anchorBoxes=[144,34;51,28; 35,33;23,25; 77,34];% by yolo
featureExtractionNetwork = resnet50;
featureLayer = 'activation_40_relu';
numClasses = 1;%width(vehicleDataset)-1;
lgraph = fasterRCNNLayers(inputSize,numClasses,anchorBoxes,featureExtractionNetwork,featureLayer);


options = trainingOptions('adam', ...  %edit with sgdm
        'MiniBatchSize', 64, ....
        'InitialLearnRate',1e-3, ...
        'MaxEpochs',50,...  % it can get good result
        'CheckpointPath', tempdir, ...       
        'ExecutionEnvironment','auto',...
        'Shuffle','every-epoch');    


doTraining=true;
 Use trainFasterRCNNObjectDetector to train Faster R-CNN object detector if doTraining is true. Otherwise, load the pretrained network.
if doTraining
    % Train the Faster R-CNN detector.
    % * Adjust NegativeOverlapRange and PositiveOverlapRange to ensure
    %   that training samples tightly overlap with ground truth.
    [detectorFast50, info] = trainFasterRCNNObjectDetector(trainingData,lgraph,options, ...
        'NegativeOverlapRange',[0 0.3], ...
        'PositiveOverlapRange',[0.6 1]);
else
    % Load pretrained detector for the example.
    pretrained = load('fasterRCNNResNet50EndToEndVehicleExample.mat');
    detectorFast50 = pretrained.detector;
end