onnxExport/onnxImport functions not working correctly

1 次查看（过去 30 天）

Luke Hubbard 2023-3-27

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1936124-onnxexport-onnximport-functions-not-working-correctly

评论： Sivylla Paraskevopoulou 2023-4-4

I'm having trouble exporting to onnx. As a test case, I tried using https://www.mathworks.com/help/deeplearning/ug/classify-sequence-data-using-lstm-networks.html. I add the following section between the training and testing blocks and got the data format error as shown below.

exportONNXNetwork doesn't raise any issues, but importONNXNetwork does.

If I import with the following option set,

importONNXNetwork('onnxEx1.onnx','OutputsDataFormat','TBC')

it works fine. All is well, right? except that onnxruntime in other environments doesn't work. Trying to load the onnx file with onnxruntime in another environment (python, java) results in a 'ShapeInferenceError' complaining that the tensor must have rank 3.

It must be that exportONNX failing to define something that is required. Are there any workarounds known for this issue? I've tried this locally with R2022b and online which uses R2023a. Both give me the same issue.

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

Sivylla Paraskevopoulou 2023-4-4

Does it help to specify the BatchSize name-value argument when you use the exportONNXNetwork function?

Also, if you are planning to use the exported network in TensorFlow, with the exportNetworkToTensorFlow function you can export directly to TensorFlow without first converting to ONNX format.

请先登录，再进行评论。

请先登录，再回答此问题。