re-train a pre-trained autoencoder

Hello, I want to retrain an autoencoder with a different set of images. autoencoder classe seems to not allowed this, so i transform my autoencoder into a classical neural network (with network function). but now i need to encode my data to train the next layer. How can i do that?

Answer 1

[Grzegorz Knor]

To encode data from retrained network you need to create new network, which contains only encoder layer. Please see the code (it is based on TrainSparseAutoencoderExample):

% load dataset
X = abalone_dataset;
% split it into two parts
X1 = X(:,1:2085);
X2 = X(:,2086:end);
% Train a first sparse autoencoder with default settings.
autoenc = trainAutoencoder(X1);
% Reconstruct inputs.
XReconstructed1 = predict(autoenc,X1);
% Compute the mean squared reconstruction error.
mseError1 = mse(X1-XReconstructed1)
% convert existed autoenc to network:
net = network(autoenc);
% retrain autoenc(net):
net = train(net,X2,X2);
% Reconstruct inputs.
XReconstructed2 = net(X2);
% Compute the mean squared reconstruction error.
mseError2 = mse(X2-XReconstructed2)
% compare biases
figure
bar([net.b{1} autoenc.EncoderBiases])
% compare weights
figure
plot(autoenc.EncoderWeights-net.IW{1})

% extract features from autoencoder
features1 = encode(autoenc,X1);

% create encoder form trained network
encoder = network;
% Define topology
encoder.numInputs = 1;
encoder.numLayers = 1;
encoder.inputConnect(1,1) = 1;
encoder.outputConnect = 1;
encoder.biasConnect = 1;

% Set values for labels
encoder.name = 'Encoder';
encoder.layers{1}.name = 'Encoder';

% Copy parameters from input network
encoder.inputs{1}.size = net.inputs{1}.size;
encoder.layers{1}.size = net.layers{1}.size;
encoder.layers{1}.transferFcn = net.layers{1}.transferFcn;
encoder.IW{1,1} = net.IW{1,1};
encoder.b{1} = net.b{1};

% Set a training function
encoder.trainFcn = net.trainFcn;

% Set the input
encoderStruct = struct(encoder);
networkStruct = struct(net);
encoderStruct.inputs{1} = networkStruct.inputs{1};
encoder = network(encoderStruct);

% extract features from net
features2 = encoder(X1);

% compare
figure
bar([features1(:,1),features2(:,1)])

If you use stacked autoencoders use encode function.

Here is a good example .

[Chris Tostado]

At the end of your post you mention "If you use stacked autoencoders use encode function." However, I'm not quite sure what you mean here. Do you mean if one were to use the "stack" function to combine multiple autoencoders, that instead of using:

% extract features from net

features2 = encoder(X1);

we should use:

% extract features from net

features2 = encode(encoder, X1); ???

The stack function, from my understanding creates a network type object in which case, the encode function no longer works. Can you please advise? Thanks!

[Giuseppe Bisazza]

I need to extract the decoder part instead of the encoder from a trained autoencoder. Could you please help me?