[bug?]Deep​LearningIm​ageClassif​icationExa​mple not getting the expected classification accuracy

Hello, I've running the DeepLearningImageClassificationExample from MATLAB.

http://uk.mathworks.com/help/vision/examples/image-category-classification-using-deep-learning.html

Unfortunately I keep getting unexpected classification accuracy. On the website the mean accuracy suppose to be 1 however the result I got was 0.3.

confMat =

         0    1.0000         0
         0    1.0000         0
    0.0213    0.9787         0

ans =

    0.3333

label =

     airplanes

I noticed for some reason the trainingFeatures are the same for all the images, after the following line:

trainingFeatures = activations(convnet, trainingSet, featureLayer, ... 'MiniBatchSize', 32, 'OutputAs', 'columns');

I'm using MATLAB 2016a and I have cuda GPU. Can someone please help to test on your machine seee if you also have the problem? Thanks a lot

Answer 1

[Joss Knight]

Do you have a new Pascal GPU like the GeForce GTX 1080? If so, this is your issue:

https://www.mathworks.com/support/bugreports/details/1439741

I do apologise for this. We are waiting on a fix from NVIDIA.

[khcy82dyc]

Yes i do! I've then tried the function on my laptop with the 980m and it worked! Thx for letting me know :)

[Joss Knight]

Patch is now published.

[Akash Maity]

Even after installing the patch, I am getting the following error while trying to run CNN classification:

Error using nnet.internal.cnngpu.convolveForward2D Unexpected error calling cuDNN: CUDNN_STATUS_BAD_PARAM. Error in nnet.internal.cnn.layer.Convolution2D/doForward (line 218) Z = nnet.internal.cnngpu.convolveForward2D( ... Error in nnet.internal.cnn.layer.Convolution2D/forwardNormal (line 195) Z = this.doForward(X,this.Weights.Value,this.Bias.Value); Error in SeriesNetwork/predict (line 139) Y(:,:,:,i) = predictNetwork.predict(X);

Is there anything that needs to be done other than installing the patch?

[Joss Knight]

Hi. I ran the entire demo on a GTX 1080 in MATLAB R2016b and got no errors. Can you describe your environment in a bit more detail? What version of MATLAB, what OS, what GPU, what are the reproduction steps? It may simply be that you have an invalid network that you trained before you installed the patch, perhaps?

[Hossam Elshahaby]

I have same problem after replacing the file cudnn64_70.dll inside the bin folder. I have Windows 10 , Matlab 2016 a , NVIDIA GEFORCE GTX , I re-trained the following net after I had replaced the file:

layers = [ imageInputLayer([40 200 1]) convolution2dLayer(10 , 96 , 'Stride',2,'Padding',0) reluLayer

    maxPooling2dLayer(2,'Stride',2)

    convolution2dLayer(3,256, 'Stride',1,'Padding',1)
    reluLayer

    maxPooling2dLayer(2,'Stride',2)

    convolution2dLayer(3,384, 'Stride',1,'Padding',1)
    reluLayer

    convolution2dLayer(3,384, 'Stride',1,'Padding',1)
    reluLayer

    convolution2dLayer(3,256, 'Stride',1,'Padding',1)
    reluLayer

    maxPooling2dLayer(2,'Stride',2)

    fullyConnectedLayer(1024)

    fullyConnectedLayer(2)
    softmaxLayer
    classificationLayer
];

[Joss Knight]

Can you upgrade MATLAB?

[khcy82dyc]

coming back to this question I'd say the best answer is to upgrade matlab...