Export LSTM to ONNX with proper input information

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.

2 个评论
显示无隐藏无

Brita Linnestad 2022-7-11

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 2022-7-12

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.

请先登录，再进行评论。