Error when using convolution2dLayer between connected maxPooling2dLayer and maxUnpooling2dLayer

Question

Bradley Treeby 2020-6-28

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/556183-error-when-using-convolution2dlayer-between-connected-maxpooling2dlayer-and-maxunpooling2dlayer

评论： Bradley Treeby 2020-6-29

I'm trying to create a modified UNet using connected max pooling and max unpooling layers. However, if I put a convolution layer between the pooling and unpooling layers, the network isn't valid. The error is reported for the unpooling layer:

Input size mismatch. Size of input to this layer is different from the expected input size.

The sizes for the CONV layer, MAXPOOL indices, and MAXPOOL size inputs are all different. Minimum working example below. Am I missing something obvious, or is it not possible to use other layers between maxpool and maxunpool?

% define layers
layers = [
    imageInputLayer([128, 128], 'Name', 'INPUTLAYER')
    maxPooling2dLayer([2 2], 'HasUnpoolingOutputs', true, 'Stride', [2 2], 'Name', 'MAXPOOL')
    convolution2dLayer([3 3], 32, 'Padding', 'same', 'Stride', [1 1], 'Name', 'CONV')
    maxUnpooling2dLayer('Name', 'UNPOOL')
    regressionLayer('Name', 'MSE')
    ]; 
% define network
lgraph = layerGraph(layers);
% define connections
lgraph = connectLayers(lgraph, 'MAXPOOL/indices', 'UNPOOL/indices');
lgraph = connectLayers(lgraph, 'MAXPOOL/size', 'UNPOOL/size');
% plot and check
analyzeNetwork(lgraph);

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Mohammad Sami 2020-6-29

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/556183-error-when-using-convolution2dlayer-between-connected-maxpooling2dlayer-and-maxunpooling2dlayer#answer_458404

在 MATLAB Online 中打开

You convolution layer is changing the number of channels in the output after the max pooling.

This causes the input size mismatch. You need to match the number of channels output by convolution layer to the output by maxpooling layer.

% define layers
layers = [
    imageInputLayer([128, 128], 'Name', 'INPUTLAYER')
    maxPooling2dLayer([2 2], 'HasUnpoolingOutputs', true, 'Stride', [2 2], 'Name', 'MAXPOOL')
    convolution2dLayer([3 3], 1, 'Padding', 'same', 'Stride', [1 1], 'Name', 'CONV')
    maxUnpooling2dLayer('Name', 'UNPOOL')
    regressionLayer('Name', 'MSE')
    ]; 
% define network
lgraph = layerGraph(layers);
% define connections
lgraph = connectLayers(lgraph, 'MAXPOOL/indices', 'UNPOOL/indices');
lgraph = connectLayers(lgraph, 'MAXPOOL/size', 'UNPOOL/size');
% plot and check
analyzeNetwork(lgraph);

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

Bradley Treeby 2020-6-29

Thanks! This pointed me to the problem in my complete code - my index to link the number of channels in the different encoding and decoding levels was offset by one.

请先登录，再进行评论。