ApplyOnnxModel with model as byte stream instead of local file

[anat-gilenson]

Is your feature request related to a problem? Please describe.
I wouldn't like to save files on my service's disk but providing a file name for an ONNX model is currently the only option when I need to load it. Instead, I would like to save the model in a remote source.

Describe the solution you'd like
I'd like to read the ONNX model from a remote source and pass it to ApplyOnnxModel as a byte stream/array instead of providing a local model file name. I see InferenceSession of OnnxRuntime is used for ONNX model loading and it provides a constructor which takes in a byte array as the model. Seems like this constructor could be used to achieve my need.

[MarcoZama]

Good feature, i'm interested!

[luisquintanilla]

Thanks for the interest in contributing @MarcoZama

@michaelgsharp when you get a chance, can you please point @MarcoZama on how to get started with his contribution. Thanks!

[matouskozak]

The new overload was added in #7254, closing as resolved.

[rosebyte]

The overload adds support for streams, but it still writes the model to a temporary file. Do we also want to support true in-memory loading, that is, loading the model directly from bytes without using temporary files?

[matouskozak]

The overload adds support for streams, but it still writes the model to a temporary file. Do we also want to support true in-memory loading, that is, loading the model directly from bytes without using temporary files?

I can see value in having it fully in-memory but out of scope of this issue. I re-opened the issue so we can discuss if we want to support the in-memory scenario.