Improve convert function

Try to read ONNX and TensorRT documents, add more conversion functions including but not limited to

• Add tensorflow model to ONNX format conversion function
• Add tensorflow saved model to tensorRT format conversion function( Note: This task has been changed) - [ ] add keras to ONNX
• Add test for these functions
• Complete convert CLI modelci modelhub convert(read our wiki)

Reference:

• https://github.com/cap-ntu/ML-Model-CI/tree/master/tests
• https://github.com/cap-ntu/ML-Model-CI/tree/master/modelci/hub/converter
• https://github.com/cap-ntu/ML-Model-CI/blob/master/modelci/hub/manager.py#L152-L195

[univerone]

I have uploaded codes about tf2onnx and tf2trt functions as well as unit testing for them. For now, all of them have run well on my machine. If there are any issues when you checked it, please inform me of the error information and i will try to fix it. Thanks!

@STALLAAAA Don’t forget to make a pull request to claim your contribution!

[dixing0908]

@STALLAAAA https://github.com/NVIDIA/sampleQAT#step-1-quantization-aware-training

Have you tried this? If not you may learn the tutorial and borrow some ideas from it.

It uses PYCUDA too. If we do need to perform test on trt inference, I still recommend to use PYCUDA anyway since it seems to be the most accepted method in community.