Multiple input (sequence and static) to a dag network

I want to train a DAG_net with two inputs, the dag network is shown below:

My two inputs are : a sequential timeseries data with 17 features for 60 training examples. The other input is a static feature for the 60 training samples. Here is a screenshot of my training data

[Yildirim Kocoglu]

Mona, have you by any chance solved this issue?

I'm trying to do the same exact thing with similar data as yours except my output is also a sequence.

I got the add/concatonate layer idea from a post I read online and a paper which shows the architecture as shown below:

My architecture looks similar to yours as below:

My training code looks something like this (to show it needs 2 inputs to train):

net = trainNetwork({XTrain(1:510,:);Static_features(1:510,:)'},YTrain(1:510,:),lgraph,options);

I use the same lgraph object as plotted above and XTrain (for temporal inputs) and Static_features (for static inputs) and a single output "YTrain" that should be a sequence similar to XTrain with multiple cells (similar to your XTrain). This is a prediction network that predicts the next time step value (YTrain) from the previous time step (XTrain).

However, when I try to train the network I get this error (next line of error says: if sequenceInputLayer is present, it can't accept any other type of input layer):

 Network: Incompatible layer types. The network contains layer types not supported with recurrent layers.
    Detected recurrent layers:
        layer 'bilstm1' (BiLSTM)
        layer 'bilstm2' (BiLSTM)
        layer 'bilstm3' (BiLSTM)
        layer 'bilstm4' (BiLSTM)
        and 1 other layers.
        

For static_input I use an imageInputLayer but, I tried every layer that exists in Matlab 2020a (apparently featureInputLayer is not avaiable in 2020a) and it still gives me the same error.

I searched for every post I believe and can't seem to find a tutorial nor an example showing how to achieve this (if it is possible). The closest thing I found was using CNN-RNN but, that is with images (frames of a video) and has a single input layer.

Answer 1

[Divya Gaddipati]

Hi,

The latest version of MATLAB allows the training of networks with multiple inputs. Please refer to the following link for more information:

https://www.mathworks.com/help/deeplearning/ug/multiple-input-and-multiple-output-networks.html

You can also use combineDatastore for this purpose.

https://www.mathworks.com/help/matlab/ref/matlab.io.datastore.combineddatastore.html

Alternatively, you can also create a custom layer that accepts multiple inputs.

https://www.mathworks.com/help/deeplearning/ug/define-custom-layer-with-multiple-inputs.html

Hope this helps!