Validation for converted models for accuracy

[varunarora]

The goal would be to get identical inference results to Paddle models and inference. @kuke has more thoughts. And here is a notebook on the topic: https://github.com/onnx/tutorials/blob/master/tutorials/CorrectnessVerificationAndPerformanceComparison.ipynb.

Q: what runtime would we use for ONNX: TensorRT?

[kuke]

A: I think mainly the server, and TensorRT is only one possible runtime environment.

[varunarora]

@kuke The server? Do you mean a Fluid ONNX backend? Because we won't have one for a while - which is why I guess you used Caffe2. Is there a reason you chose Caffe2?

[kuke]

Yes, recently many users show interests in converting the models from Fluid to ONNX, then run with other backends on PC or mobile devices.

I used Caffe2 because the built-in backend in ONNX doesn't work, here is the issue.