How to Solve this, Semantic Segmentation with Dice Loss

3 次查看（过去 30 天）

显示更早的评论

Mohammad Bhat 2019-12-18

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/497201-how-to-solve-this-semantic-segmentation-with-dice-loss

评论： Mohammad Bhat 2019-12-21

在 MATLAB Online 中打开

My images are with 256 X 256 in size

I am doing semantic segmentation with dice loss.

ds = pixelLabelImageDatastore(imdsTrain,pxdsTrain);
layers = [
         imageInputLayer([256 256 1])
         convolution2dLayer(3,32,'Padding',1)
         reluLayer
         maxPooling2dLayer(2,'Stride',2)
         convolution2dLayer(3,32,'Padding',1)
         reluLayer
         transposedConv2dLayer(4,32,'Stride',2,'Cropping',1)
         reluLayer
         transposedConv2dLayer(4,32,'Stride',2,'Cropping',1)
         convolution2dLayer(1,3)
         softmaxLayer
         dicePixelClassificationLayer('dice')
         ]
opts = trainingOptions('sgdm',  'InitialLearnRate',1e-3,  'MaxEpochs',10,  'LearnRateDropFactor',1e-1,  'LearnRateDropPeriod',50,  'LearnRateSchedule','piecewise',  'MiniBatchSize',32);
net = trainNetwork(ds,layers,opts);

Errors are :-

Error using trainNetwork (line 165)

Invalid training data. The output size ([512 512 3]) of the last layer does not match the response size ([256 256 3]).
Error in net_net (line 110)
net = trainNetwork(ds,layers,opts);
The output size ([512 512 3]) of the last layer does not match the response size ([256 256 3]).

4 个评论
显示 2更早的评论隐藏 2更早的评论

Walter Roberson 2019-12-18

I do not know what the cause of the problem is, but you could try adding a resize layer on the bottom.

Mohammad Bhat 2019-12-21

在 MATLAB Online 中打开

I solved the problem rather error by adding maxpooling layer at the end....

.......................................................................

......................................................................

layers = [
    imageInputLayer([256 256 1])
    convolution2dLayer(3,32,'Padding',1)
    reluLayer
    maxPooling2dLayer(2,'Stride',2)
    convolution2dLayer(3,32,'Padding',1)
    reluLayer
    transposedConv2dLayer(4,32,'Stride',2,'Cropping',1)
    reluLayer
    transposedConv2dLayer(4,32,'Stride',2,'Cropping',1)
    reluLayer
    maxPooling2dLayer(2,'Stride',2)
    convolution2dLayer(1,3)
    softmaxLayer
    dicePixelClassificationLayer('dice')
    ]