How to set the encoder transfer function in autoencoder

I want to set the encoder transfer function by myself. But I do not how to do it. Can the encoder transfer function be changed by myself

Answer 1

[SANA]

First you make an autoencoder and generate its function, the code for generating the function of autoencoder is:

autoenc = trainAutoencoder(Data, 300,...

'MaxEpochs', 100,...

'L2WeightRegularization', 0.001,...

'SparsityRegularization', 4,...

'SparsityProportion', 0.05,...

'ScaleData', true);

generateFunction(autoenc)

The generated function open in MATLAB editor with the name of neural_function, I renamed it my_autoencoder and the transfer function is mentioned there, so you can edit it as you wish, code is below:

function [y1] = my_encoder(x1)

%NEURAL_FUNCTION neural network simulation function.

%

% Generated by Neural Network Toolbox function genFunction, 15-Nov-2018 15:50:26.

%

% [y1] = neural_function(x1) takes these arguments:

% x = 5088xQ matrix, input #1

% and returns:

% y = 5088xQ matrix, output #1

% where Q is the number of samples.

%#ok<*RPMT0>

% ===== NEURAL NETWORK CONSTANTS =====

% Input 1

x1_step1.xoffset = [-10;-11;-11;-12;-1]

x1_step1.gain = [0.090;0.9090;0.9090;0.90909;0.0769]

x1_step1.ymin = 0;

% Layer 1

b1 = [-0.115;0.768;0.7066;0.5009303396;0.1249019302]

IW1_1 = [-0.0947 0.7320 0.3146 0.494636173 -0.0951171]

b2 = [3.5409901027442902688;3.6635759144424437928]

LW2_1 = [-1.1707436273955371675 1.5786236406994880177 -1]

% Output 1

y1_step1.ymin = 0;

y1_step1.gain = [0.0909090909090909;0.07]

y1_step1.xoffset = [-10;-11;-11;-12;-12]

% ===== SIMULATION ========

% Dimensions

Q = size(x1,2); % samples

% Input 1

xp1 = mapminmax_apply(x1,x1_step1);

% Layer 1

a1 = logsig_apply(repmat(b1,1,Q) + IW1_1*xp1);

% Layer 2

a2 = logsig_apply(repmat(b2,1,Q) + LW2_1*a1);

% Output 1

y1 = mapminmax_reverse(a2,y1_step1);

end

% ===== MODULE FUNCTIONS ========

% Map Minimum and Maximum Input Processing Function

function y = mapminmax_apply(x,settings)

y = bsxfun(@minus,x,settings.xoffset);

y = bsxfun(@times,y,settings.gain);

y = bsxfun(@plus,y,settings.ymin);

end

% **********************************************

% ************ Enhance encoder here ************

% Sigmoid Positive Transfer Function

function a = logsig_apply(n,~)

a = 1 ./ (1 + exp(-n));

end

% ************ Enhance encoder here ************

% **********************************************

% Map Minimum and Maximum Output Reverse-Processing Function

function x = mapminmax_reverse(y,settings)

x = bsxfun(@minus,y,settings.ymin);

x = bsxfun(@rdivide,x,settings.gain);

x = bsxfun(@plus,x,settings.xoffset);

end

You can change decoder function as well. Enjoy !!!

[huilin zhu]

Thanks very much.But do you mean I have to edit a autoencoder before I edit the transfer function, and then I have to train my autoencoder again?

Answer 2

[Frédéric BERTHOMMIER]

I changed the decoder transfer function to recover a PCA equivalent which I checked::

autoenc = trainAutoencoder(Data,4,'MaxEpochs',5000,'DecoderTransferFunction','purelin');

This substitutes to the generation of a .m function as proposed before for obtaining a PCA equivalent. Note that the encoder transfer function cannot be modified.