Alphapose Onnx and TensorRT models

[mrn-mln]

Hi, I've been working on converting Alphapose (the model which contributors have developed for X86 devices at #113) to onnx and tensorRT.
I will send a PR after finishing it. here I wanna share my experience with this procedure.
I tried to convert a Pytorch-based model of HRnet net (one of the available backbones of Alphapose) to Onnx but the path was not straightforward to me.
First of all, I use Pytorch version 1.1.0 for exporting the onnx model but the version was buggy and I faced several errors. after trying some versions of PyTorch finally the below versions worked for me.

`
torch==1.5.1
torchvision==0.6.1

//Expporting parameters
pose_model.load_state_dict(torch.load(MODEL.pt, map_location=map_location))
dummy_input = torch.randn(1, 3, 256, 192, requires_grad=True).cuda()
torch.onnx.export(pose_model, dummy_input, "alphapose.onnx", export_params=True, opset_version=11)
`
Next step, I will check the onnx model results and compare them to the x86 model result in order to make sure that the onnx model is exported successfully.

I will update the procedure here.
please feel free to ask for more details if you need them.

[mrn-mln]

Alphapose estimator is converted to TRT and tested on jetson-tx2 (10-12 fps on jetson-tx2 with 2 people on average at each frame)
The source code is available here.

[christian-holland]

I will update the procedure here. please feel free to ask for more details if you need them.

Hey sounds great, do you have some more description or a script for the conversion produce?

[mrn-mln]

I will update the procedure here. please feel free to ask for more details if you need them.

Hey sounds great, do you have some more description or a script for the conversion produce?

You can check Neuralet website for more information.
https://docs.neuralet.com/en/latest/
https://neuralet.com/products/edge-adaptive-learning/