The resulting weights and biases of the autoencoder training doesn't give the same encoded value when calculated manually.

clc

clear all

close all

%% Generate mxk Cosine wave training data with noise

n = [0:99]';

b = 0.15*pi;

w = b*ones(1,1000);

S = cos(w.*n);

[m,k] = size(S);

a = 0.375; % noise variance

X = S+sqrt(a)*randn(m,k);

%% Define the no. of hidden layers

hiddenSize = 2;

%% Build the autoencoder

autoenc = trainAutoencoder(X,hiddenSize,...

'MaxEpochs', 1000, ...

'EncoderTransferFunction','satlin',...

'DecoderTransferFunction','purelin',...

'L2WeightRegularization',0.01,...

'SparsityRegularization',4,...

'SparsityProportion',0.10);

%% View AE structure

view(autoenc)

%% Generate 10 testing data and plot against predicted and original signals

for i=1:10

X_test = cos(b*n)+sqrt(a)*randn(m,1);

xReconstructed = predict(autoenc,X_test);

subplot(2,1,1), plot(n, X_test,n, xReconstructed,'Linewidth',2)

grid on

legend('Noisy','Reconstructed')

%% ..........................

subplot(2,1,2), plot(n, cos(b*n),n, xReconstructed,'Linewidth',2)

grid on

legend('Original','Reconstructed')

pause

end

WW = autoenc.EncoderWeights; % WW are the weights after autoencoder training

BB = autoenc.EncoderBiases; % BB are the biases after autoencoder training

X_test = cos(b*n);

encoded_result = encode(autoenc,X_test)

encoded_result =

0.0956 % the value of the hidden neurons after using the function encoded

0.0859

then we use the equation of encoded which is

z=h(1)(W(1)x+b(1)),

our code is

encoded_manually = ((WW * X_test) + BB);

we use sigmoid function

for i3 = 1:hiddenSize

logsig(i3,1) = 1 / (1 + exp(-encoded_manually(i3,1)));

end

logsig

finally our results :

logsig =

0.6631

0.4902

So, there is a mismatch between the results when calculated using the function "encode" and when calculated manually using the function.provided in the website for the function "encode"

if anyone can help us to know where is the mistake in our code and how can we have same result between encoded function and manual function

Answer 1

[Anshika Chaurasia]

Hi,

The reason for discrepancy is that the inputs are preprocessed before encoding i.e., inputs are transformed using the 'mapminmax' function.

You can also notice this by converting the autoencoder to a network object:

net = network(autoenc);
net.inputs{1}.processFcns

Try the following snippet to manually calculate:

net = network(autoenc);
% preprocess data
for iii = 1:numel(net.inputs{1}.processFcns)
    inputs = feval( net.inputs{1}.processFcns{iii}, ...
        'apply', X_test, net.inputs{1}.processSettings{iii} );
end
% manually encode 
encoded_manually = satlin((WW * inputs) + BB)