My simple numerical Deep Learning project will not train. Get "invalid training data" message.

Hello Matlab,

I am trying to use Deep Learning GUI to process 5 numerical inputs per training trial to match one numerical goal. Total of 6 entries per line.

I had this working in Excel but can't go beyond 200 variables for their solver.

My training data in Excel is a row for each trial with my "goal or correct answer" in the right column as Matlab wants.

I bring the array into matlab as a "matrix". It is around 100 rows deep, 6 across.

From there, I drag this matrix into the arrayDatastore function, then say

>> bb = arrayDatastore(athurvaluesforDL,"ReadSize", 6,"IterationDimension", 2,"OutputType", "cell")

I have tried both 1 and 2 as InterationDimension.

The simple network has an input, 2 fully connecteds separated by a relu and the regression output.

The architecture checks out as at least "viable".

It won't train. I get this message:

Training with trainNetwork failed. Invalid training data. For a network with 1 inputs and 1 output, the datastore read function must return a cell array with 2 columns, bit it returns a cell array with 1 columns.

So I am missing something. I am new to Matlab and have spent about 3 days getting to know the screen basics.

I went through the help page on arrayDatastore, but maybe I need something else.

Most of the "help" on Deep Learning is about imaging. I just want to do numerics.

Thanks.

Thursday 6 16 22 The question is withdrawn. I no longer have the Matlab DL kit.

Answer 1

[Ben]

I think the issue here is you need to permute the data to NumFeatures x BatchSize for featureInputLayer when working with trainNetwork and datastores. Here's an example:

% Create fake data with 100 observations of 5 features and 1 target

xy = randn(100,6);

% split off features from targets

x = xy(:,1:5);

y = xy(:,6);

% permute for featureInputLayer

% For featureInputLayer and trainNetwork using datastores the data has to

% be shaped as NumFeatures x BatchSize

x = x.';

y = y.';

% Create datastores with cell outputs, and iterate along dimension 2 (since

% we permuted the data it is now in NumFeatures x BatchSize).

xds = arrayDatastore(x,"OutputType","cell","IterationDimension",2);

yds = arrayDatastore(y,"OutputType","cell","IterationDimension",2);

% Combine the datastores

cds = combine(xds,yds);

% Create a simple network and train.

layers = [featureInputLayer(5)

fullyConnectedLayer(1)

regressionLayer];

opts = trainingOptions("adam");

net = trainNetwork(cds,layers,opts);

[Yongwon Jang]

xds = arrayDatastore(x,"OutputType","cell","IterationDimension",2);
yds = arrayDatastore(y,"OutputType","cell","IterationDimension",2);

cds = combine(xds,yds);

Create arrayDatastore X, Y separately and combine the two d into one.

I couldn't find any example code containing this content anywhere.

My problem was putting input arguments in the form of an arrayDatastore into the minibatchqueue.

I found this answer by chance and it helped me a lot.

Thank you.