How to use "imageInputLayer" instead of "sequenceInputLayer"?

I want to build a layergraph for deep learning with multiple sequence inputs. Since matlab does not seem to support multiple sequenceinputlayers (https://de.mathworks.com/matlabcentral/answers/709528-why-does-multiple-inputs-with-sequenceinputlayer-return-an-error) I am using "imageInputLayer" instead of "sequenceInputLayer".

Since the error I get is:

Error using trainNetwork

Invalid training data. Predictors and responses must have the same number of observations.

I wanted to try a minimal example, and switching out the input layers dos not work here either:

lgraph = [...
    %sequenceInputLayer(12,"Name","sequence_2")
    imageInputLayer( [12 1] , "Name", "sequence_2", "Normalization", "none")
    tanhLayer("Name","tanh_2")
    fullyConnectedLayer(1,"Name","fc_2",'WeightsInitializer','he')
    softmaxLayer("Name","softmax_2")
    regressionLayer("Name","regressionoutput")];

It works with the sequenceInputLayer, but not with the imageInputLayer. The error code is:

Error using trainNetwork

Invalid training data. The output size ([1 1 1]) of the last layer does not match the response size ([1 1 63100]).

The training data is 12x63100 and the output is 1x63100, and since the fullyConnectedLayer should output 1x63100, I do not understand why the imageInputLayer does not work here.

Answer 1

[Ben]

Your imageInputLayer([12,1]) is specifying that your input data is "images" with height 12, width 1 and 1 channel/feature.

I expect that since your data is 12x63100 then there are 2 potential cases:

1. You have 63100 sequences of length 12 with 1 feature.
2. You have a single sequence of length 63100 with 12 features.

In case 1. you would typically create imageInputLayer([12,1,1]) and permute your data so that it has size 12x1x1x63100. This allows you to create and train networks that use convolution2dLayer like a 1D convolution over sequences with fixed length. Similarly in case 2. you can use imageInputLayer([63100,1,12]) and permute your data to have size 63100x1x12. In both cases you can technically use featureInputLayer too.

A more modern way to handle networks with multiple sequence inputs would be to use dlnetwork and a custom training loop following examples such as this one.