BentoML

I see, that's interesting. It doesn't make much sense in this case, does it? I mean ONNX makes model independent from Python package and that the whole point of exporting?

Not necessarily, the point of ONNX is that you can use a framework that’s optimizied for training to train the model, e.g. PyTorch, and uses another framework that’s optimized for inferencing/serving to load the same model, e.g. Tensorflow, Caffe2, TensorRT

We are investigating a more connivent way to use ONNX model with BentoML, without the user to load the model before packing with a TensorflowModelArtifact for example.  It might be something like this:
```@artifact([OnnxModelArtifact("model", backend="Tensorflow"]))
class MyService(BentoService):
    pass

svc = MyService()
svc.pack('model', './path_to_model_file.onnx')
svc.save()```
