Make onnx dependencies optional

[jppgks]

hummingbird-ml depends on onnxconverter-common. onnxconverter-common 1.12 introduced a dependency on onnxruntime. We had to pin onnxconverter-common: ludwig-ai/ludwig#2347.

At Ludwig, we do not use the onnx features from hummingbird, so we'd like to avoid having to install them.

[ksaur]

We use onnxconverter-common in our internal IR. I'm not sure why onnxconverter-common 1.12 introduced a dependency on onnxruntime (microsoft/onnxruntime#11640).

@xadupre Do you have any suggestions?

As a temporary fix, since we already have "onnxmltools>=1.6.0,<=1.11.0" in the optional deps due to #592, we could also make it "onnxconverter-common>=1.6.0,<=1.11.0" to avoid inadvertantly taking a dependency on onnxruntime

[xadupre]

I'll look into that. This is a surprise to me. There is no mention of onnxruntime in onnxconverter-common on github.

[mshr-h]

HB 0.4.4 depends on onnxmltools>=1.6.0. The onnxmltools==1.12.0 tries to import onnxconverter-common.auto_mixed_precision.py which depends on onnxruntime.
I think that's why ...

As mentioned by @ksaur, we've temporarily fixed the issue with #592. But we haven't released the stable release with it.

@ksaur @interesaaat
I think it's a good time to release v0.4.5. What do you think?

[xiaowuhu]

update onnxconverter-common to 1.12.1 to fix this issue, please have a try.