Problems with layers dimensions in CNN

Asked Feb 24 '18 at 16:10

Active Feb 27 '18 at 16:59

Viewed 96 times

I took this neural network configuration from https://arxiv.org/pdf/1603.01068.pdf and I do not understand why Matlab says that layer 8 cannot accept the data from layer 7 due to dimensions problems. These are the layers:

layers = [
    imageInputLayer([block_dimensions block_dimensions 3]) %64x64x3

    convolution2dLayer([4 4],32,'NumChannels',3,'Stride',1)  %61x61x32
    maxPooling2dLayer([2 2],'Stride',2)  %31x31x32

    convolution2dLayer([5 5],48,'NumChannels',32,'Stride',1)  %27x27x48
    maxPooling2dLayer([2 2],'Stride',2)  %14x14x48

    convolution2dLayer([5 5],64,'NumChannels','auto','Stride',1)  %10x10x64
    maxPooling2dLayer([2 2],'Stride',2)  %5x5x64

    convolution2dLayer([5 5],128,'NumChannels','auto','Stride',1)  %1x1x128
    fullyConnectedLayer(128)
    reluLayer()
    fullyConnectedLayer(cameras_number)
    softmaxLayer()
    classificationLayer
];

I've calculated all the sizes of layers output and they seem to be right. If I remove the last maxpooling layer, the network works but it performs not well as I expect.

edited Feb 27 '18 at 16:59

asked Feb 24 '18 at 16:10

Luca Di Liello

1,486
2
17
34

What is block size in your input layer? Is it possible you are reducing tensor size below 1 in Width or Height by the successive pooling ops? – Alex Taylor Feb 24 '18 at 18:09
Input blocks have size 64x64x3 – Luca Di Liello Feb 24 '18 at 18:56
Do you know how I can retrieve the dimensions I calculated from the network? – Luca Di Liello Feb 24 '18 at 18:58

Problems with layers dimensions in CNN

0 Answers0