Preparing data for regression using deep neural network

Hi,

I'm trying to implement a deep neural network for regression with hand-crafted features as the network input. I'm trying to use the Deep Network Designer to achieve this. The network archiecture is similar to the diagram below:

• Each input/feature is a scalar array of length 14751, and there are 9 inputs/features alogether,
• there is one output, again a scalar array of length 14751,
• and there are 4 samples altogether.
• See the data attached.

load data.mat;

inputSize = length(inputs)

inputSize = 9

[~, numSamples] = size(output) % where each column represents a different sample

numSamples = 4

exampleInput = inputs{1,1}; 
size(exampleInput)

ans = 1×2

14751 4

Can someone please advise how I can go about preparing the raw data in Datastore format which can be loaded in to Deep Network Designer?

[David Ho]

Hello OB,

The best way to import your data into Deep Network Designer by forming two array datastores and combining into a single datastore using the combine function.

For the data you provided, you can do the following:

load data.mat

numFeatures = size(inputs,1);

% All inputs have the same number of timesteps so we
% can get the sizes by looking at the first observation
numTimesteps = size(inputs{1}, 1);
numObservations = size(inputs{1}, 2);

% Reshape the inputs from a cell to an array
arrayInputs = zeros(numFeatures, numTimesteps, numObservations);
for ii = 1:numFeatures
    arrayInputs(ii,:,:) = inputs{ii};
end

% Generate ArrayDatastores for the inputs and outputs
adsPredictors = arrayDatastore(arrayInputs, "IterationDimension", 3);
adsResponses = arrayDatastore(output', "IterationDimension", 1);

% Generate a combined datastore for the inputs and outputs together
cds = combine(adsPredictors, adsResponses);

This datastore is suitable for use with a network that has a sequenceInputLayer expecting data with 9 features, and a regressionOutputLayer outputting sequences with one feature.

For more information on creating datastores for Deep Network Desginer, you can refer to the following resources:

https://uk.mathworks.com/help/deeplearning/ug/import-data-into-deep-network-designer.html

https://uk.mathworks.com/help/deeplearning/ug/datastores-for-deep-learning.html

[OB]

Hello David,

Thanks a lot for your response. I now seem to be able to load the data. However, am still struggling with training the network. Here is the architecture I'm trying to start with:

• 9 input features (as previously described)
• 2 fully-connected hidden layers with 32 neurons each
• a sigmoid activation functions
• 1 ouput

Here is the code from Deep Network Designer:

layers = [
    sequenceInputLayer(9,"Name","sequence")
    sigmoidLayer("Name","sigmoid_1")
    fullyConnectedLayer(32,"Name","fc_1")
    sigmoidLayer("Name","sigmoid_2")
    fullyConnectedLayer(32,"Name","fc_2")
    sigmoidLayer("Name","sigmoid_3")
    regressionLayer("Name","regressionoutput")];

For training data, I'm loading the combined datastore (cds) as per your steps.

However, when trying to train, I get the following error:

Could you please advise?

[Milan Bansal]

Change the output size of the last fully connected layer to 1.