LSTM architecture for a sequence-to-sequence model

I have build a LSTM sequence-to-sequence model which I would like to convert to onnx and use in a python script-

My input data is as follows, 11307 sequences of 24 times steps and 6 features:

XTrain = 11307x1 cell array

{ 6x24 double}

YTrain = 11307x1 cell array

{1x24 catgorical}

The layer is :

numFeatures = 6

NumHiddenUnis = 50

numClasses = 2

layers = [ ...

sequenceInputLayer(numFeatures)

lstmLayer(numHiddenUnits,'OutputMode','sequence')

fullyConnectedLayer(numClasses)

softmaxLayer

classificationLayer];

netlstm=trainNetwork(XTrain,YTrain,layers,options)

Whe I run the model on test data with numFeatures = 6, sequence_length = 24 in matlab, the results are good.

Exporting netlstm to onnx using

exportONNXNetwork(netlstm, 'file.onnx')

Viewing the file.onnx in a model checker show that the initial_h is 1x1x50, not 1x24x50 as expected.

Loading the onnx.model in python, I can run the model with input 1x1x6, but not he full sequence of 24 time steps.

Why is the sequnce length missing from initial_h after exporting to onnx?

The example: Sequence-to-Sequence Classification Using Deep Learning - MATLAB & Simulink - MathWorks Nordic

gives the same issue when exporting the network after running the example.

Answer 1

[Sivylla Paraskevopoulou]

You can try specifying the BatchSize name-value argument of the exportONNXNetwork function.