Why a great software such as Matlab does not support some ONNX operators and where and when can I find mfiles for replacing the placeholder layers?

Even Matlab R2019b does not support, some simple ONNX operators such as ‘Gather’, ‘shape’, ‘Resahpe’ and so on. As a results, the Matlab users have been encountered a serious problems for importing the new ONNX models. It is not good for Matlab as a great software. Does anyone can help me on this issue or is there any mfile for replacing unsupported ONNX operators?

[Sarah Mohamed]

Hi Hossein,

We don't currently support all ONNX operators, but we're working on supporting more models in the future. If you can share some more information about which models you are trying to import, and where they are exported from, that would be very helpful.

If you are using 'importONNXLayers', any unsupported operator will be replaced by a PlaceholderLayer in the imported layer graph. There are a few functions that can be used to remove/replace PlaceholderLayers, but you may need to know how the unsupported operators are being used in the ONNX file in order to do this.

• findPlaceholderLayers can be used to locate all of the unsupported operators, which are imported as PlaceholderLayers in the layer graph
• replaceLayer can be used to replace a PlaceholderLayer with a new layer. If you can't replace the PlaceholderLayer with one of the supported deep learning layers, you can write your own custom layer. Here is a documentation page that describes how to write a custom layer: https://www.mathworks.com/help/deeplearning/ug/define-custom-deep-learning-layers.html
• removeLayers can be used to delete an unnecessary layer from a layer graph

If you'd like some more help, could you share your ONNX file(s)?

Thanks,

Sarah