MATLAB Answers

TrainNetwork 4 dimension error

6 views (last 30 days)
yaspy caroline
yaspy caroline on 24 Oct 2019
Answered: jibrahim on 25 Oct 2019
My project is Source separation using masking. Here is my code below where I get an error at trainNetwork as "Invalid validation data. X must be a 4-D array of images.", even after converting the data into 4D array. Could someone help me?
clc;clear all;close all;
%Training data
[y,fs] =audioread('13069 tri.wav');
[z,fs] =audioread('7040 sine.wav');
ny=length(y);
nz=length(z);
N=min([ny nz]);
s1=y(1:N);
s2=z(1:N);
left = s1(:,1);
right = s1(:,1);
time = (1/fs)*length(s1);
t = linspace(0, time, length(left));
plot(t,left, t, right);
xlabel('time(sec)');
ylabel('signal strength');
% sound(left, fs);
% sound(right, fs);
% sound(d, fs);
% l= left(10:15);
% r=right(10:15);
% source1= s1(10:15,:);
left = s2(:,1);
right = s2(:,1);
time = (1/fs)*length(s2);
t = linspace(0, time, length(left));
plot(t,left, t, right);
xlabel('time(sec)');
ylabel('signal strength');
% sound(left, fs);
% sound(right, fs);
% sound(d, fs);
% l= left(10:15);
% r=right(10:15);
% source2= s2(11:15,:);
S1=reshape(s1,[],4);
S2=reshape(s2,[],4);
%Testing data
[y1,fs] =audioread('13069 tri.wav');
[z1,fs] =audioread('7040 sine.wav');
ny1=length(y1);
nz1=length(z1);
N=min([ny1 nz1]);
ss1=y1(1:N);
ss2=z1(1:N);
left = ss1(:,1);
right = ss1(:,1);
time = (1/fs)*length(ss1);
t = linspace(0, time, length(left));
plot(t,left, t, right);
xlabel('time(sec)');
ylabel('signal strength');
% sound(left, fs);
% sound(right, fs);
% sound(d, fs);
% l= left(10:15);
% r=right(10:15);
%ssource1= ss1(10:15,:);
left = ss2(:,1);
right = ss2(:,1);
time = (1/fs)*length(ss2);
t = linspace(0, time, length(left));
plot(t,left, t, right);
xlabel('time(sec)');
ylabel('signal strength');
% sound(left, fs);
% sound(right, fs);
% sound(d, fs);
% l= left(10:15);
% r=right(10:15);
%ssource2= ss2(11:15,:);
S11=reshape(s1,[],4);
S22=reshape(s1,[],4);
%Mix
mixTrain = S1 + S2;
mixTrain = mixTrain / max(mixTrain);
mixValidate = S11 + S22;
mixValidate = mixValidate / max(mixValidate);
%cwt train
P_mix0 = cwt(mixTrain);
P_M = abs(cwt(s1));
P_F = abs(cwt(s2));
%test cwt
P_Val_mix0 = cwt(mixValidate);
P_Val_M = abs(cwt(ss1));
P_Val_F = abs(cwt(ss2));
maskTrain = P_M ./ (P_M + P_F + eps);
%Compute the validation soft mask. Use this mask to evaluate the mask emitted by the trained network.
maskValidate = P_Val_M ./ (P_Val_M + P_Val_F + eps);
X=reshape(P_mix0,[],4);
Y=reshape(maskTrain,[],4);
P_Val_mix=reshape(P_Val_mix0,[],4);
maskValidate1=reshape(maskValidate,[],4);
[X] = digitTrain4DArrayData;
[Y] = digitTrain4DArrayData;
P_Val_mix=digitTrain4DArrayData;
maskValidate1=digitTrain4DArrayData;
idx = randperm(size(X,4),1000);
X1= X(:,:,:,idx);
X(:,:,:,idx) = [];
idx1 = randperm(size(Y,4),100);
Y1= Y(:,:,:,idx1);
Y(:,:,:,idx1) = [];
newY=zeros(size(X));
idx2 = randperm(size(P_Val_mix,4),1000);
P_Val_mix1= P_Val_mix(:,:,:,idx2);
P_Val_mix(:,:,:,idx2) = [];
idx3 = randperm(size(maskValidate1,4),100);
maskValidate11= maskValidate1(:,:,:,idx3);
maskValidate1(:,:,:,idx3) = [];
newY=zeros(size(X));
% layers =[
% imageInputLayer([28 28 1])
% convolution2dLayer([3,1],1,'Padding','same')
% reluLayer
% convolution2dLayer([3,1],1,'Padding','same')
% reluLayer
% convolution2dLayer([3,1],1,'Padding','same')
% reluLayer
% convolution2dLayer([3,1],1,'Padding','same')
% reluLayer
% regressionLayer
% ];
layers = [ ...
imageInputLayer([28 28 1],"Normalization","None")
layers =[
imageInputLayer([28 28 1])
convolutionLayer([3,1],1,'Padding','same')
reluLayer
convolutionLayer([3,1],1,'Padding','same')
reluLayer
convolutionLayer([3,1],1,'Padding','same')
reluLayer
convolutionLayer([3,1],1,'Padding','same')
reluLayer
regressionLayer
];
maxEpochs = 3;
miniBatchSize = 28;
options = trainingOptions("adam", ...
"MaxEpochs",maxEpochs, ...
"MiniBatchSize",miniBatchSize, ...
"SequenceLength","longest", ...
"Shuffle","every-epoch",...
"Verbose",0, ...
"Plots","training-progress",...
"ValidationFrequency",30,...
"ValidationData",{P_Val_mix0,maskValidate},...
"LearnRateSchedule","piecewise",...
"LearnRateDropFactor",0.9, ...
"LearnRateDropPeriod",1);
%Do training
CocktailPartyNet = trainNetwork(X,newY,layers,options);
%CocktailPartyNet = trainNetwork(mixSequencesT,maskSequencesT,layers,options);
estimatedMasks0 = predict(CocktailPartyNet,P_Val_mix);
estimatedMasks0 = estimatedMasks0.';
Softs1Mask = estimatedMasks0;
Softs2Mask = 1 - Softs1Mask;
P_Val_mix0 = P_Val_mix0(:,1:size(Softs1Mask,2));
P_s1 = P_Val_mix0 .* Softs1Mask;
P_s1 = [conj(P_s1(end-1:-1:2,:)) ; P_s1];
source1_est_soft = icwt(P_s1);
source1_est_soft = source1_est_soft / max(abs(source1_est_soft));
% range = (numel(win):numel(maleSpeech_est_soft)-numel(win));
t = (1/fs);
figure(9)
subplot(2,1,1)
plot(t,S11)
title("Original Source1")
xlabel("Time (s)")
grid on
subplot(2,1,2)
plot(t,source1_est_soft)
xlabel("Time (s)")
title("Estimated source 1 (Soft Mask)")
grid on
% sound(maleSpeech_est_soft(range),Fs)
%Multiply the mix STFT by the female soft mask to get the estimated female speech STFT. Use the ISTFT to get the estimated male audio signal. Scale the audio.
P_s2 = P_Val_mix0 .* Softs2Mask;
P_s2 = [conj(P_s2(end-1:-1:2,:)) ; P_s2];
source2_est_soft = icwt(P_s2);
source2_est_soft = source2_est_soft / max(source2_est_soft);
%Visualize the estimated and original female signals. Listen to the estimated female speech.
figure(10)
subplot(2,1,1)
plot(t,S22)
title("Original Source 2")
grid on
subplot(2,1,2)
plot(t,source2_est_soft)
xlabel("Time (s)")
title("Estimated source 2 (Soft Mask)")
grid on

  0 Comments

Sign in to comment.

Answers (1)

jibrahim
jibrahim on 25 Oct 2019
Hi Yaspy,
The problem is in the dimensions of your validation data: P_Val_mix0,maskValidate
They're not 4D

  0 Comments

Sign in to comment.

Sign in to answer this question.

Products


Translated by