How to train dataset after dimension reduced with autoencoder into a support vector machine?

I am doing a anomaly detection by using stacked autoencoder and one-class svm. The stacked autoencoder is done. The problem is I am unsure of how to use the dimension reduced data after the autoencoder to be trained in the one-class svm?

Here's the code for the stacked autoencoder:

train1 = fullfile(dataFolder, "train_FD001.txt");

[train_data1, train_labels1] = importdata(train1)

% To avoid results from training are different each time

% Explicitly set the random number generator seed

rng('default')

% Specifying the values for the regularizers for the training

%autoenc1 = trainAutoencoder(train_data1,hiddenSize1, ...

autoenc1 = trainAutoencoder(train_data1,100,'MaxEpochs',400,'L2WeightRegularization',1.0, ...

'DecoderTransferFunction','purelin','UseGPU',true)

% Visualize the first autoencoder

view(autoenc1)

% Train the next autoencoder on a set of these vectors extracted from the training data.

% First, you must use the encoder from the trained autoencoder to generate the features.

feat1 = encode(autoenc1,train_data1)

% Specifying the values for the regularizers for the training

autoenc2 = trainAutoencoder(feat1,10,'MaxEpochs',400,...

'DecoderTransferFunction','purelin','UseGPU',true)

% Visualize the second autoencoder

view(autoenc2)

% Train the next autoencoder on a set of these vectors extracted from the training data.

% First, you must use the encoder from the trained autoencoder to generate the features.

feat2 = encode(autoenc2,feat1)

% Specifying the values for the regularizers for the training

autoenc3 = trainAutoencoder(feat2,1,'MaxEpochs',400,...

'DecoderTransferFunction','purelin','UseGPU',true)

% Visualize the third autoencoder

view(autoenc3)

% Train the next autoencoder on a set of these vectors extracted from the training data.

% First, you must use the encoder from the trained autoencoder to generate the features.

feat3 = encode(autoenc3,feat2)

%% Training the final softmax layer

% Train a softmax layer to classify the 50-dimensional feature vectors.

% Unlike the autoencoders, you train the softmax layer in a supervised fashion using labels for the training data.

softnet = trainSoftmaxLayer(feat3,train_data1,'MaxEpochs',400)

% Vizualize the diagram of the softmax layer

view(softnet)

%% Forming a stacked neural network

stackednet = stack(autoenc1,autoenc2,autoenc3,softnet)

% Vizualize the stacked autoencoder

view(stackednet)

Answer 1

[yanqi liu]

yes，sir，may be use the encoder net to get data low dimension vector as feature，reshape them to feature matrix and label vector，then train it in new svm model

[Ng Yong Jie]

May I know how can I reshape the feature matrix and label vector?

[yanqi liu]

yes，sir，because not see your data，just as use autoencoder before，use origin data by encoder model get low dimension vector，and save it into matrix

fea_vec = encode(your_encoder_model, origin_data);

[Ng Yong Jie]

It results in an error.

I have trained each individual autoencoder and stacked them up into a stacked autoencoder. I believe now my problem is to put my data into the stacked autoencoder for the feature extraction. Then, use the extracted data and put it into the one-class SVM for anomaly detection.

[yanqi liu]

no，sir，should use every model to get feature vector，not use stack net

[Ng Yong Jie]

For that, I believe I've done it in the code.

feat1 = encode(autoenc1,train_data1)

feat2 = encode(autoenc2,feat1)

feat3 = encode(autoenc3,feat2)

[Ng Yong Jie]

So, is it to make feat3 into matrix for the SVM anomaly detection?

[yanqi liu]

yes，sir，may be use feat3 as input，of course，its dimension reduce，may be consider use feat1、feat2

[Ng Yong Jie]

I tried doing that but it doesn't work.