Description
I'm finding that ONNX Runtime actually has very good performance compared to TFLite in the situations I'm using it, and I'm becoming inclined towards migrating to ONNX. However, ONNX model conversion from TF seems to be missing a key feature: support for multiple signatures.
As far as I can see, if a model has multiple signatures you are forced to select one of them using the --signature_def argument. However, in my case I need to be able to export and run multiple different functions that run in a same model. Creating separate models for each signature would be quite wasteful, as I have many functions and many of them share weights.
Would it be possible to add support for multiple signatures? If not, is this a limitation of TF to ONNX conversion, or a limitation of ONNX itself?
By the way, TensorFlow Lite has support for multiple signatures too since version 2.7, and --signature_def is currently limited to saved models. Though that's not what I'm asking here, maybe you want to consider extending this argument to work on TFLite models too.