How to build TensowFlow on Jetson TK1

I have posted all the steps required to build TensorFlow on Jetson TK1 at:

It takes a long time to build and it is really pushing the Jetson TK1 to its limits, but it can be done.

I have also updated the wheel file at https://drive.google.com/file/d/0B1uGKNpQ7xNqZ2pvSmc3SlZJS2c/view?usp=sharing
if you are just interested in installing it. The build requires CUDNN 6.5 but it is using the standard 6.5 toolkit.

Enjoy TensorFlow.

Awesome, Thanks for the post.

I’m using your instructions to try and build this for my TX1 and it is going pretty well.
I’ll do a post with the couple differences from your instructions when I get done.

I do have one question though. I have the most recent jetpack installed on my board.
It has cuda 7.0 but also only has libcudnn 4.0.2. Where can one find libcudnn 6.5?

I don’t see it here https://developer.nvidia.com/embedded/downloads

Thanks

The libcudnn 6.5 is at https://developer.nvidia.com/rdp/cudnn-archive ( look for v2).

That is great.
One question. Is it possible to cross-compile Tensorflow and run the application on TK1.
I wondered if it would be possible to use Google Tango or Nexus 9 as the host.
Can you elaborate on this please.

I know you can cross-compile from x86-64 ( one of my colleagues has done it).

Thank you sir, you’re a gentleman and a scholar! I’ve been waiting for this =)

I’ll try to get it up and running this weekend and report back.