How to use C/C++ with Tensorflow on Jetson TX2

Robotics & Edge Computing Jetson & Embedded Systems Jetson TX2
1

Is there any method to do inference with Tensorflow C/C++ api on Jetson TX2? I use the method in this post [url]https://devtalk.nvidia.com/default/topic/1038957/jetson-tx2/tensorflow-for-jetson-tx2-/[/url] to install Tensorflow on TX2, but there is no Tensorflow c_api.h file can be found.

2

Hi, yes. I’ve the same concern. Is there any way to run custom tf models on tx2? Layers in my model aren’t yet supported by TensorRT so I can’t use its c++ api either. How do I run an almost realtime inference on tx2 without TensorRT? Is there any official or even unofficial way to do inference with Tensorflow C/C++ api on Jetson TX2?

3

Hi,

Our official TensorFlow package is only built with python interface.
Some users wants to build TensorFlow C++ library on Jetson but meet some AWS issue:
[url]https://devtalk.nvidia.com/default/topic/1044430/jetson-agx-xavier/how-to-build-the-tensorflow_cc-so-tensorflow-c-interface-on-the-xavier-/[/url]

AFAIK, there is no TensorFlow C++ libraries available on Jetson now.
A recommended way is to inference your model with TensorRT and implement the non-supported layer with our plugin API.
[url]https://docs.nvidia.com/deeplearning/sdk/tensorrt-developer-guide/index.html#plugin_sample[/url]

Thanks.

4

So when will there be support for C++ libraries on the TX2?

1 Like
5

Any updates on this? I’m trying to compile tensorflow 1.13.1 from source on the TX2 and am having issues with Eigen.

Eigen/src/Core/arch/GPU/PacketMath.h(152): error: calling a __device__ function ("__int_as_float") from __host__ __device__ function("eq_mask") is not allowed


Eigen/src/Core/arch/GPU/PacketMath.h(152): error: calling a __device__ function ("__longlong_as_double") from __host__ __device__ function("eq_mask") is not allowed


output 'tensorflow/core/kernels/_objs/inplace_ops_gpu/inplace_ops_functor_gpu.cu.pic.o' was not created
6

I didn’t find the solution to develop with C/C++ APIs of Tensorflow on Jetson TX2. Actually, I tend to use MXNet on Jetson TX2, and it works well so far.