Custom CNN layer Dot Indexing Issue?

1 次查看（过去 30 天）

Matt 2021-9-6

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1447819-custom-cnn-layer-dot-indexing-issue

评论： Matt 2021-9-8

Hello!

I'm working on a CNN and I would like to be able to connect the output of a fully-connected layer to a convolutional layer. From some other "ask" pages, I read that a custom reshape layer is the solution for this. I have tried implementing one and I'm having some issues. When I use the checkLayer function, Matlab says everything is working. And when I insert the layer into my architecture and check it, Matlab doesn't throw any errors. However, when I try to train my CNN, I get the following error:

{Error using trainNetwork (line 184)

Dot indexing is not supported for variables of this type.

Error in trainDiscriminator (line 34)

[Disc_2, round1Data]=trainNetwork(trainData, trainLabels, Disc_2, trainOpts);

Caused by:

Dot indexing is not supported for variables of this type.

}

There's not really a way to determine if that's the custom layers' fault, but deleting them from the design did fix the error. I've also tested several similar architectures without the custom layers and none of them threw an error. For reference, the code for my reshape layer is as follows:

classdef reshapeLayer < nnet.layer.Layer
    properties
        rows;
        cols;
    end
    properties (Learnable)
    end
    methods
        function layer = reshapeLayer(name,rows,cols)
            layer.Name = name;
            layer.rows=rows;
            layer.cols=cols;
        end
        function [Z] = predict(layer,X)
            Z = reshape(X,layer.rows,layer.cols,1,[]);
        end
    end
end

Does anyone know how I can address this issue regarding the dot indexing of custom layers?

2 个评论
显示无隐藏无

Abolfazl Chaman Motlagh 2021-9-7

在 MATLAB Online 中打开

Hi, i test your code. and it runs without error. i put your layer and test it on simple Cifar10 dataset.

mds = imageDatastore('images\cifar10Train\','LabelSource',"foldernames",'IncludeSubfolders',true,'ReadFcn',@(x) im2gray(imread(x)));
layers = [imageInputLayer([32 32 1],'Normalization',"none");
reshapeLayer('test',32,32);
fullyConnectedLayer(10);
softmaxLayer;
classificationLayer]
net = trainNetwork(imds,layers,options);