C++ Interface for model deployment on zcu104 platform instead of python

Questions
1

Hi @jtuyls/@mak ,

We are trying to deploy FR on the ZCU104 platform, where I have generated shared libraries for detector and as well feature extract models. To build an end to end FR application we also need pre/post processing functions along with model inference shared libs.

And implementation of the pre/post processing modules are available with us in form of C++ code.

Could you please help me to guide how to deploy C++ application on zcu104 platforms using TVM shared library files.

Thanks and Regards, Raju

2

Hi @jtuyls / @mak ,

Is there any software components dependency? like pre-built SDCard package files Do we have a image read and display tools/libraries with this pre-built SDCard package ? like OpenCV

If not, please help me to direct some reference link where we can build and integrate those tools and if any.

Thanks and Regards, Raju

3

@kvaju.454

We don’t have cpp application for cloud and Edge. You can follow this example for cpp application.

4

@kvaju.454 The instructions for setting up the software can be found here: Vitis-AI Integration — tvm 0.8.dev0 documentation. There is a dependency on pyxir, h5py and pydot. Is this what you mean? Also, as pointed to by @mak you can use the usual TVM C++ APIs for deployment.

5

Hi @mak / @jtuyls ,

Thanks for your quick reply on this. I will go through these links and get back to you if needed.

Thanks and Regards, Raju

6

Hi @jtuyls,

There is a dependency on pyxir, h5py and pydot. Is this what you mean?

→ Do we need these packages to load and run the tvm_dpu_arm.so model files?

If yes, then we can’t use the C++ interface right, is my understanding correct?

Thanks and Regards, Raju

7

@kvaju.454

Pyxir build requires h5py. However, this would not effect your cpp interface.

8

Hi @mak ,

Thanks for the information, I am working on it and I will be updating further on this.

Thanks and Regards, Raju