Deep Learning Toolbox Multiply by [] in Second Learning Iteration?

Question

Travis Wiens 2022-7-5

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1754135-deep-learning-toolbox-multiply-by-in-second-learning-iteration

回答： Abhaya 2024-9-5

In the following code for a Conditional Variational Autoencoder I get the following error on the second iteration. I don't see any unusual values in the learnables for the encoderNet or decoderNet. It appears that this error doesn't appear when I run it in 2022a using the "run" button on this page.

Output in 2021a:

>> minimum_cvae

iteration =

1

loss =

1(C) × 1(B) single dlarray

6.6875

iteration =

2

loss =

1(C) × 1(B) single dlarray

6.6573

Error using .*

Arrays have incompatible sizes for this operation.

Error in tp26c2fafb_59b5_412b_af58_81dab1a42f5d (line 36)

tmp_33 = ((tmp_30.*[]).*constants{2});

Error in

deep.internal.recording.convert.tapeToFunction>@(varargin)fcnWithConstantsInput(varargin{:},constants)

(line 37)

fcn = @(varargin)fcnWithConstantsInput(varargin{:},constants);

Error in deep.internal.recording.CodegenExtensionMethod/backward (line

61)

[varargout{1:meth.NumGradients}] =

meth.BackwardFunction(varargin{:});

Error in deep.internal.dlarray.ExtensionOperation/backward (line 46)

[varargout{1:nargout}] = backward(op.Method, varargin{:});

Error in deep.internal.recording.RecordingArray/backwardPass (line 70)

grad = backwardTape(tm,{y},{initialAdjoint},x,retainData,false);

Error in dlarray/dlgradient (line 83)

[grad,isTracedGrad] = backwardPass(y,xc,pvpairs{:});

Error in minimum_cvae>modelGradients (line 64)

[decodeGrad, encodeGrad] = dlgradient(loss, decoderNet.Learnables, ...

Error in deep.internal.dlfeval (line 18)

[varargout{1:nout}] = fun(x{:});

Error in dlfeval (line 41)

[varargout{1:nout}] = deep.internal.dlfeval(fun,varargin{:});

Error in minimum_cvae (line 47)

[encodeGrad, decodeGrad] = dlfeval(...

Related documentation

Code:

%generate some fake data

XTrain=dlarray(randn(10,100),'CB');%Autoencoder inputs

YTrain=dlarray(randn(2,100),'CB');%Conditional inputs

N_input=size(XTrain,1);

N_latent = 9;%number of latent variables

N_cond=size(YTrain,1);%number of conditional inputs

%define encoder with two inputs

encoderLG = layerGraph([

featureInputLayer(N_input,'Name','input_encoder');

concatenationLayer(1,2,'Name','concat1')

fullyConnectedLayer(5,'Name','fc1')

reluLayer('Name','relu1')

fullyConnectedLayer(2 * N_latent, 'Name', 'fc_encoder')

]);

encoderLG = addLayers(encoderLG,featureInputLayer(N_cond,'Name','cond_input'));

encoderLG = connectLayers(encoderLG,"cond_input","concat1/in2");

%definte decoder

decoderLG = layerGraph([

featureInputLayer(N_latent,'Name','input_decoder');

concatenationLayer(1,2,'Name','concat1')

fullyConnectedLayer(5,'Name','fc1')

reluLayer('Name','relu1')

fullyConnectedLayer(N_input, 'Name', 'fc_decoder')

]);

decoderLG = addLayers(decoderLG,featureInputLayer(N_cond,'Name','cond_input'));

decoderLG = connectLayers(decoderLG,"cond_input","concat1/in2");

encoderNet = dlnetwork(encoderLG);

decoderNet = dlnetwork(decoderLG);

numEpochs = 10;

lr = 1e-3;

iteration=0;

avgGradientsEncoder = [];

avgGradientsSquaredEncoder = [];

avgGradientsDecoder = [];

avgGradientsSquaredDecoder = [];

for epoch = 1:numEpochs

iteration=iteration+1

[encodeGrad, decodeGrad] = dlfeval(...

@modelGradients, encoderNet, decoderNet, XTrain, YTrain);

[decoderNet.Learnables, avgGradientsDecoder, avgGradientsSquaredDecoder] = ...

adamupdate(decoderNet.Learnables, ...

decodeGrad, avgGradientsDecoder, avgGradientsSquaredDecoder, iteration, lr);

[encoderNet.Learnables, avgGradientsEncoder, avgGradientsSquaredEncoder] = ...

adamupdate(encoderNet.Learnables, ...

encodeGrad, avgGradientsEncoder, avgGradientsSquaredEncoder, iteration, lr);

end

iteration = 1

loss =

1(C) × 1(B) single dlarray 6.5279

iteration = 2

loss =

1(C) × 1(B) single dlarray 6.5519

iteration = 3

loss =

1(C) × 1(B) single dlarray 6.5803

iteration = 4

loss =

1(C) × 1(B) single dlarray 6.5597

iteration = 5

loss =

1(C) × 1(B) single dlarray 6.5038

iteration = 6

loss =

1(C) × 1(B) single dlarray 6.5244

iteration = 7

loss =

1(C) × 1(B) single dlarray 6.4807

iteration = 8

loss =

1(C) × 1(B) single dlarray 6.5267

iteration = 9

loss =

1(C) × 1(B) single dlarray 6.4935

iteration = 10

loss =

1(C) × 1(B) single dlarray 6.4696

function [encodeGrad, decodeGrad] = modelGradients(encoderNet, decoderNet, x, y)

[z, zMean, zLogvar] = sampling(encoderNet, x, y);

xPred = sigmoid(forward(decoderNet, z, y));

loss = ELBOloss(x, xPred, zMean, zLogvar)

[decodeGrad, encodeGrad] = dlgradient(loss, decoderNet.Learnables, ...

encoderNet.Learnables);

end

function elbo = ELBOloss(x, xPred, zMean, zLogvar)

squares = 0.5*(xPred-x).^2;

weight=10;

reconstructionLoss = mean(squares(:));

KL = mean( -.5 * mean(1 + zLogvar - zMean.^2 - exp(zLogvar), 1));

elbo = weight*reconstructionLoss + KL;

end

function [zSampled, zMean, zLogvar] = sampling(encoderNet, x, y)

encoded = forward(encoderNet, x, y);

d = size(encoded,1)/2;

zMean = encoded(1:d,:);

zLogvar = encoded(1+d:end,:);

sz = size(zMean);

epsilon = randn(sz);

sigma = exp(.5 * zLogvar);

z = epsilon .* sigma + zMean;

zSampled = dlarray(z, 'CB');

end

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

Abhaya 2024-9-5

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1754135-deep-learning-toolbox-multiply-by-in-second-learning-iteration#answer_1511344

Hello Travis,

I encountered a similar error while using the Deep Learning Toolbox in MATLAB 2021a.

This error appears to occur during a call to the ‘backwardPass’ function. However, when I run the same code in MATLAB 2021b, it works without any issues for me.

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。

Deep Learning Toolbox Multiply by [] in Second Learning Iteration?

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

回答（1 个）

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

另请参阅

类别

标签

产品

版本

Community Treasure Hunt

Deep Learning Toolbox Multiply by [] in Second Learning Iteration?

0 个评论 显示 -2更早的评论隐藏 -2更早的评论

回答（1 个）

0 个评论 显示 -2更早的评论隐藏 -2更早的评论

另请参阅

类别

标签

产品

版本

Community Treasure Hunt

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

0 个评论
显示 -2更早的评论隐藏 -2更早的评论