Currently, our MXNET-ONNX export module does not support dynamic input shapes. We have to specify the static input shapes and the mxnet shape inference api will deduce the static output shapes. I propose to add dynamic input shapes support with a minimalist option dynamic_input_shapes=True/False to our contrib.onnx.export_model api. This option will override the mxnet shape inference and allow for fully or partially dynamic input shapes e.g. (None, None) or (None, 32).

Dynamic shapes might reduce the level of ONNX graph optimization. Specific performance comparisons between static and dynamic shapes can be further investigated and documented (on some standard models).

The success criterion of this feature is that all models that previously can be exported and run with static shapes should also work with dynamic shapes.

It would be great if the dev community can provide some comments or specific use cases for this feature.