Invalid Training data. The Output size of the last layer does not match the reponse size

18 次查看（过去 30 天）
Zehao Pan
Zehao Pan 2021-7-27
评论： Zehao Pan 2021-7-30
So I'm getting this error on one computer while on another I don't.
Invalid training data. The output size ([32 32 2]) of the last layer does not match the response size ([32 32 1]).
So his program takes an image and segment out a part of it. The image and the mask that represent the correct segemention are all [32 32 1]
The result it spits out is [32 32 2] because its like the % chance of it being the thing to segment out and then the % chance its the background which is why it has 2 dimensions instead of 1 at the end. I just don't understand why it runs fine on one computer but have this error on another computer when its exact same code. The only difference is that it works on computer using 2020b and its not working on the computer 2021a.
Edit: I tested with 2020b on both and the second computer still get the same error while the first computer doesn't.
imageDir = fullfile('..\Data_Processing\trainingImages');
%location of mask
labelDir = fullfile('..\Data_Processing\trainingLabels');
imds = imageDatastore(imageDir);
classNames = ["fluid", "background",];
labelIDs = [254 1];
pxds = pixelLabelDatastore(labelDir,classNames,labelIDs);
augmenter = imageDataAugmenter('RandRotation',[0 90],'RandXReflection',true);
%create batches for training data
patchSize = [32 32];
patchPerImage = 160;
miniBatchSize = 32;
patchds = randomPatchExtractionDatastore(imds,pxds,patchSize, ...
'PatchesPerImage',patchPerImage,'DataAugmentation',augmenter);
patchds.MiniBatchSize = miniBatchSize;
minibatch = preview(patchds);
disp(minibatch);
%create unet
imageSize = [32 32];
numClasses = 2;
lgraph = unetLayers(imageSize, numClasses);
options = trainingOptions('sgdm', ...
'InitialLearnRate',1e-3, ...
'MaxEpochs',5, ...
'VerboseFrequency',10);
predictPatchSize = [32 32];
[net,info] = trainNetwork(patchds,lgraph,options);

请先登录，再回答此问题。

采纳的回答

KSSV
KSSV 2021-7-28
This is due to number of filters at Final-convolution layer. DEfault it has number of filters as 2, you need to change it to 1. You can do the following:
  1. Type deepNetworkDesigner() at the command window.
  2. A gui named deep netowrk desinger wqill pop out. (be patient wait for sometime)
  3. Once gui s opneed, import the created lagraph into gui and you can see your network architecture.
  4. USe, zoomin, view,fit to width options and go to the Final Convolution layer. Thirs one from the last.
  5. Click on that block, on the right you can see it's properties.
  6. At the NumFilters it is 2, change it to 1
  7. Export the network to workspace, it will be saved as lgraph_1.
  8. Now you try training using the exported lgraph_1.
[net,info] = trainNetwork(patchds,lgraph_1,options);
Note all the above steps, can also be achieved via command lines; you need to read them at the doc.
  2 个评论
Zehao Pan
Zehao Pan 2021-7-28
编辑：Zehao Pan 2021-7-28
So this does allow it to run but my mini batch results are just 100% now which doesn't seem right?
The part that I'm still confused about is why does it work fine on one computer but another it give this error? For my original computer it runs and spit back out like 32 32 2 results. one graph behing like % chance each location is the mask and the other graph being % chance each location is the background. I use basically just the first graph to recreate the mask (and give decent results on some)
Running a smaller data to check the network but giving me just 100% on all the mini batches feels like its wrong? Will update once I can check it.
Edit: Ya it just like mark everything as the mask instead.
Zehao Pan
Zehao Pan 2021-7-30
I figured out the error which was somethind really stupid. I had open the code on a older version of matlab first which give me error with classNames = ["fluid", "background",]; making me change a ll the " " to ' '...
So this made it so my classname was just 1 string fluidbackground instead of 2 and that's why the error was appearing.

请先登录，再进行评论。

更多回答（0 个）

类别

在 Help Center 和 File Exchange 中查找有关 Image Data Workflows 的更多信息

标签

产品


版本

R2020b

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by