Export LSTM to ONNX with proper input information

I have created a LSTM network and converted it to onnx using matlabs exportONNXNetwork. The onnx-network will be loaded in Java using OrtSession.

Layers: [5x1 nnet.cnn.layer.Layer]

layers =

Sequence Input Sequence input with 6 dimensions (numberOfFeatures)

LSTM LSTM with 50 hidden units

Fully connected 2 fully connected layers

Softmax softmax

Clasification Output crossentropyex

Sequence length is 24.

Using exportONNXNetwork(netlstm,filename), the only reported input is 'sequenceinput'.

How can i set up exportONNXNetwork so the onnx-model holds more/all input information needed when loading the model in Java?

Answer 1

[Sivylla Paraskevopoulou]

I am not sure what you mean by "more/all input information". If you mean that you want a network that can be used for prediction, you must train the layer graph that you created and then export the trained network and not the layer graph.

[Brita Linnestad]

I have trained the layer graph, and then exported the trained network.

When loading the trained network in Java using OrtSession, I get an ortsession runtimeerror :

Non-zero status code returned while running LSTM node. Name:'lstm' Status Message: Input initial_h must have shape {1,24,50}. Actual:{1,1,50}

How can I, before I export my model from Matlab, set initial_h or other information needed for OrtSession to run properly?

[Sivylla Paraskevopoulou]

In MATLAB, if your input data is a vector sequence, the sequenceInputLayer expects the data in the format CSN, where C is the number of features or channels, S is the sequence length, and N is the number of observations. For an example on how to train a network with a vector sequence input, Train Network for Sequence Classification.

When you export the network to ONNX, the input tensor shape should be NSC. I am not sure what is happenning to the input when you convert from ONNX to ortSession.