Invalid training data. Sequence responses must have the same sequence length as the corresponding predictors.

Hi , I have a simple regression NN for time sequences .

Got an error when training the network : "Invalid training data. Sequence responses must have the same sequence length as the corresponding predictors. " . While my XTrain data is a cell array dim(Nx1) of sequences dim(1x10) and YTrain data is cell array dim(Nx1) of sequences sim(1x1) ...Please advise why the response sequences need to be the same length as the predictors ?

Architecture of the network :

Answer 1

[yanqi liu]

yes，sir，may be set sequenceInputLayer to

sequenceInputLayer(10, ……)

to match the input data，1*10 dimension

[Igor Lisogurski]

Hi , while changing what you adviced i am getting the error below

[yanqi liu]

yes，sir，may be upload the data mat file，we can debug

[Igor Lisogurski]

Hi , i attached the neaural network creation and training flow , as well the data set ...You need only to change the path of the "load()" function in the script to the data set path in your computer .

Thanks a lot for help

[yanqi liu]

close all; clc;
%% load data
data=load('tx2rx_connection_rx_pow_minus10dBm__21_12_2021.mat');
x_org=data.in_tx;
y_org=data.tx2rx;
%% normalize data
x=((real(x_org)-mean(real(x_org)))/std(real(x_org))) +1j*(imag(x_org)-mean(imag(x_org)))/std(imag(x_org));
y=((real(y_org)-mean(real(y_org)))/std(real(y_org))) +1j*(imag(y_org)-mean(imag(y_org)))/std(imag(y_org));
%% figure #1 - Input/Output Time Domain
figure
ax1=subplot(2,1,1);plot(real(x));hold on;plot(real(y));hold off;
title('Time Domain - Real Parts');ylabel('[V]');xlabel('[sec]');legend('x','y');

ax2=subplot(2,1,2);plot(imag(x));hold on;plot(imag(y));hold off;
title('Imaginary Parts');xlabel('[sec]');ylabel('[V]');legend('x','y');

linkaxes([ax1,ax2],'xy');
%% figure #2 -Input/Output Frequency Domain
figure

%FFT PSD
% nfft=1024;
% psd_x_real=fftshift(fft(real(x),nfft)).*conj(fftshift(fft(real(x),nfft)))/(length(x)*fs);
% psd_x_imag=fftshift(fft(imag(x),nfft)).*conj(fftshift(fft(imag(x),nfft)))/(length(x)*fs);
% psd_y_real=fftshift(fft(real(y),nfft)).*conj(fftshift(fft(real(y),nfft)))/(length(x)*fs);
% psd_y_imag=fftshift(fft(imag(y),nfft)).*conj(fftshift(fft(imag(y),nfft)))/(length(x)*fs);

%Welch PSD
segment_size=1024;
nfft=2048;
noverlap=segment_size-2;
fs=128e6;
[pwelch_x,f_x]=pwelch(real(x),segment_size,[],nfft,fs,'centered','power');
[pwelch_y,f_y]=pwelch(real(y),segment_size,[],nfft,fs,'centered','power');

ax3=subplot(2,1,1);plot(f_x,10*log10(pwelch_x/1e-3));hold on;plot(f_y,10*log10(pwelch_y/1e-3));hold off;
title('Power Spectrum');xlabel('[Hz]');ylabel('[dBm]');legend('x','y')

ax4=subplot(2,1,2);plot(f_x,10*log10(pwelch_x/1e-3));hold on;plot(f_y,10*log10(pwelch_y/1e-3));hold off
xlabel('[Hz]');ylabel('[dBm]');ylim([-5,15]);legend('x','y')

grid on;

%% NN data preparation %%
train_ratio=0.9;
channel_length=5;
feature_dim=1;

%Divide into train/test arrays
x_train_len=floor(train_ratio*length(x));
y_train_len=floor(train_ratio*length(y));

x_train=[zeros(1,channel_length-1),(x(1:x_train_len))'];
x_test=[zeros(1,channel_length-1),(x(x_train_len+1:end))'];
y_train=y(1:y_train_len);
y_test=y(y_train_len+1:end);

% NN training/validation inputs definition
XTrain=cell(length(y_train),1);
XTest=cell(length(y_test),1);
YTrain=cell(length(y_train),1);
YTest=cell(length(y_test),1);

x_train_sample=zeros(1,2*channel_length);
y_train_sample=zeros(1,1);

XTrain = [];
YTrain = [];
for i=1:length(y_train)
    x_train_sample(1,1:channel_length)=real(x_train(i:i+channel_length-1));
    x_train_sample(1,channel_length+1:end)=imag(x_train(i:i+channel_length-1));
    
    y_train_sample(1,1)=real(y_train(i));
    
    XTrain(:,:,1,i)= x_train_sample(:);
    YTrain(i,1)= y_train_sample;
end

x_test_sample=zeros(1,2*channel_length);
y_test_sample=zeros(1,1);

XTest = [];
YTest = [];
for i=1:length(y_test)
    x_test_sample(1,1:channel_length)=real(x_test(i:i+channel_length-1));
    x_test_sample(1,channel_length+1:end)=imag(x_test(i:i+channel_length-1));
    
    y_test_sample(1,1)=real(y_test(i));
    
    XTest(:,:,1,i)= x_test_sample(:);
    YTest(i,1)= y_test_sample;
end
%% NN Architecture %%
hidden_neuron=8;

layers = [ ...
    imageInputLayer([10 1 1],'Name','Sequence Input Layer')
    convolution2dLayer(3,hidden_neuron,'Padding','same','Name','CNN 2d Layer')
    reluLayer('Name','ReluLayer Hidden Layer')
    fullyConnectedLayer(1,'Name','CNN 2d Output Layer')
    reluLayer('Name','ReluLayer Output Layer')
    regressionLayer('Name','Regression NN')];

lgraph = layerGraph(layers);
analyzeNetwork(lgraph)

epochs=50;
learnign_r=0.0005;
batch_size=128;
drop_out_r=0;
options = trainingOptions('adam', ...
    'MaxEpochs',epochs, ...
    'ValidationData',{XTest,YTest}, ...
    'InitialLearnRate',0.0005, ...
    'MiniBatchSize',32, ...
    'Plots','training-progress',...
    'LearnRateDropFactor',drop_out_r);
%% NN Training %%
[net,info] = trainNetwork(XTrain,YTrain,layers,options);

[Igor Lisogurski]

Hi @yanqi liu sorry for late reply ....Thanks a lot , it worked for me , appriciate your help

[yanqi liu]

yes，sir，it is my pleasure