Custom Vnet layer validation failed error

I am trying to train a custom Vnet network it keeps giving me the following error

Layer 'en3_prelu1': Layer validation failed. Error using

'predict' in Layer prelu3dLayer. The function threw an error

and could not be executed.

Array dimensions must match for binary array op.

This is my class definition

classdef prelu3dLayer < nnet.layer.Layer

% Custom 3-D PReLU layer.

%

% Copyright 2019 The MathWorks, Inc.

properties (Learnable)

% Layer learnable parameters

% Scaling coefficient

Alpha

end

methods

function layer = prelu3dLayer(numChannels, name)

% layer = preluLayer(numChannels, name) creates a PReLU layer

% with numChannels channels and specifies the layer name.

% Set layer name.

layer.Name = name;

% Set layer description.

layer.Description = "PReLU with " + num2str(numChannels) + " channels";

% Initialize scaling coefficient.

layer.Alpha = rand(numChannels);

end

function Z = predict(layer, X)

% Z = predict(layer, X) forwards the input data X through the

% layer and outputs the result Z.

Z = max(0, X) + layer.Alpha .* min(0, X);

end

function [dLdX, dLdAlpha] = backward(layer, X, ~, dLdZ, memory)

% [dLdX, dLdAlpha] = backward(layer, X, Z, dLdZ, memory)

% backward propagates the derivative of the loss function

% through the layer.

% Inputs:

% layer - Layer to backward propagate through

% X - Input data

% Z - Output of layer forward function

% dLdZ - Gradient propagated from the deeper layer

% memory - Memory value which can be used in backward

% propagation

% Outputs:

% dLdX - Derivative of the loss with respect to the

% input data

% dLdAlpha - Derivative of the loss with respect to the

% learnable parameter Alpha

dLdX = layer.Alpha .* dLdZ;

dLdX(X>0) = dLdZ(X>0);

dLdAlpha = min(0,X) .* dLdZ;

%dLdAlpha = sum(sum(dLdAlpha,1),2);

% Sum over all observations in mini-batch.

dLdAlpha = sum(dLdAlpha,5);

end

end

end

And these are the network options

inputSize = [40 40 40];

numClasses = 2;

lgraph = createVnetBn(inputSize, numClasses);

figure;

lgraph.plot

maxEpochs = 250;

options = trainingOptions('adam', ...

'MaxEpochs',maxEpochs, ...

'InitialLearnRate',1e-3, ...

'LearnRateSchedule','piecewise', ...

'LearnRateDropPeriod',5, ...

'LearnRateDropFactor',0.97, ...

'ValidationData',dsVal, ...

'ValidationFrequency',400, ...

'Plots','training-progress', ...

'Verbose',false, ...

'MiniBatchSize',miniBatchSize);

doTraining = true;

if doTraining

modelDateTime = datestr(now,'dd-mmm-yyyy-HH-MM-SS');

[net,info] = trainNetwork(dsTrain,lgraph,options);

save(['trained3DUNet-' modelDateTime '-Epoch-' num2str(maxEpochs) '.mat'],'net');

end

Answer 1

[Amanjit Dulai]

It's tricky to tell without the code for createVnetBn, but I would guess the problem is this line in prelu3dLayer:

% Initialize scaling coefficient

layer.Alpha = rand(numChannels);

This will initialize layer.Alpha to a numChannels-by-numChannels matrix, which is not what you would normally want for a 3D input. For a 3D input, you probably want something like this:

% Initialize scaling coefficient

layer.Alpha = rand([1 1 1 numChannels]);

[Amira Youssef]

It didn't solve the problem

[Amanjit Dulai]

Is there any way to provide the code for createVnetBN?

[Amira Youssef]

https://www.mathworks.com/matlabcentral/fileexchange/71521-3-d-deep-learning-lung-tumor-segmentation

the original code is here but he is using 64x64x64 mat files while I am using 40x40x40

[Amanjit Dulai]

I got a different error when running the code:

Error using rand

Size inputs must be integers.

Error in prelu3dLayer (line 25)

layer.Alpha = rand(numChannels);

The problem here is that V-nets and U-nets have a particular architecture where sizes are divided by 2 through the depth of the network (see the diagram on page 3: https://arxiv.org/pdf/1606.04797.pdf). For this reason, only some input sizes are valid for a V-net or U-net. Normally sizes that are powers of 2 (like 64) will work well. A size of 40x40x40 doesn't work in this instance because 40/(2*2*2*2) = 2.5.

You could try using unet3dLayers, which lets you control the depth of the network, and so will let you reduce the number of downsampling layers so an input size of 40x40x40 will work. Or, you can edit createVnetBn to remove the 5th downsampling stage and the 1st upsampling stage.

[Amira Youssef]

I tried to fix the createVnet but it gives unconnected layers and doesn't run as well, do you have any idea what exactly should be edited

[Amanjit Dulai]

I've attached an edited version of createVnetBn.m which should work for an input size of 40x40x40. I have removed one of the downsampling and upsampling stages.

[Amira Youssef]

Thank you so much sir, I really can't thank you enough for this