TensorFlow 2.0 cross-compilation for TX2

Hello,

the latest release of TensorFlow wheel file is 1.14 which is not the latest major version of this framework (https://developer.download.nvidia.com/compute/redist/jp/v42/tensorflow-gpu/).

I am wondering is there any tutorial for cross-compiling TF 2.0 for Jetson TX2 with CUDA 10.0? It seemed to work when built on the device, but it’s not very efficient as it takes long time.

Thank you.

Hi,

Here is an official tutorial for compiling TensorFlow for ARM environment:
https://www.tensorflow.org/lite/guide/build_arm64

By the way, here is a prebuilt TensorFlow v2.0 package contributed by the community.
https://devtalk.nvidia.com/default/topic/1065109/jetson-nano/building-tensorflow-2-0-and-tensorflow-lite-python-runtime/

Thanks.

Hi,

thank you, but in case I want to modify and compile not lite but general version of TF? Is it possible?

AFAIK https://developer.download.nvidia.com/compute/redist/jp/v42/tensorflow-gpu/ is providing general version, which I want, but it’s an older release - 1.14. My goal is to cross-compile latest release - 2.0. Same as NVIDIA does with 1.14 internally.

Thanks

Hi,

The steps should be similar for the general TensorFlow.

Thanks.

Hi,
I am trying to cross-compile the general TensorFlow-2 for my TX2, but could not find a similar build_aarch64_lib.sh file for the general version in the source code nor a CROSSTOOL file for supporting it. Could you elaborate on how to cross-compile the general version?
Thanks.

Hi yxy171230,

Please help to open a new topic. Thanks