TensorFlow on Jetson TX2

Here’s my pip wheel: https://drive.google.com/file/d/0Bxl-G9VJ61mBYmZPY0hLSlFaUDg/view?usp=sharing
And here is the step by step installation process: https://syed-ahmed.gitbooks.io/nvidia-jetson-tx2-recipes/content/first-question.html

Do you delete the swap file after installation? Thx!

does TensorFLow has GPU support?

Yes I did delete the swap file after installation. After you reboot, the swapfile is not used and hence is safe to delete. If you have an SSD, it might be better to keep the swapfile in use in the SSD as mentioned in JetsonHacks: http://www.jetsonhacks.com/2014/10/04/creating-swapfile-ubuntu-nvidia-jetson-tk1/

Yes it does. The pip wheel I referred here is with the GPU support.

Thank you. Very helpful guide. For anyone who is planning to do this, I would definitely recommend adding the SSD, as the regular build of TF takes a lot of time.

Thank you tousif111! That helped a lot!

Here’s a pip wheel for Python 3.5 and TensorFlow R1.1 :
https://drive.google.com/file/d/0B-FNNHo8QQ5LSmc1R0JkSm5fVUE/view?usp=sharing

belosthomas. thanks
I always failed to build the version 1.1
how did you do that?

I followed the tutorial sent by tousif111: https://syed-ahmed.gitbooks.io/nvidia-jetson-tx2-recipes/content/first-question.html.

I used Bazel 0.5.0 which include the CPU.java fix :
https://github.com/bazelbuild/bazel/releases/download/0.5.0/bazel-0.5.0-dist.zip

May be the version of the bazel was the problem with me
I will try the upper version of it
thanks

any luck with version 1.2.1 ?
Upgrade completed.
Seems functioning

Just finished building the wheel for TF 1.2.1 on Jetson TX2:

tensorflow-1.2.1-cp27-cp27mu-linux_aarch64.whl

using bazel 0.5.2 + patch for eigen stuff + API switch
Minimal test (‘Hello World!’) runs

tensorflow-1.2.1-cp27-cp27mu-linux_aarch64.whl (and md5 signature) uploaded:

https://drive.google.com/drive/folders/0B1I3Wk04wyI5Qk9ReU1mUGVtbk0?usp=sharing

md5sum -b tensorflow-1.2.1-cp27-cp27mu-linux_aarch64.whl > tensorflow-1.2.1-cp27-cp27mu-linux_aarch64.whl.md5 md5sum -c tensorflow-1.2.1-cp27-cp27mu-linux_aarch64.whl.md5
tensorflow-1.2.1-cp27-cp27mu-linux_aarch64.whl: OK

tensorflow-1.2.1-cp27-cp27mu-linux_aarch64.whl (and md5 signature) uploaded:

https://drive.google.com/drive/folders/0B1I3Wk04wyI5Qk9ReU1mUGVtbk0?usp=sharing

md5sum -b tensorflow-1.2.1-cp27-cp27mu-linux_aarch64.whl > tensorflow-1.2.1-cp27-cp27mu-linux_aarch64.whl.md5 md5sum -c tensorflow-1.2.1-cp27-cp27mu-linux_aarch64.whl.md5
tensorflow-1.2.1-cp27-cp27mu-linux_aarch64.whl: OK

Hello,

I followed the installation process, but now the Jetson tx2 does not load the Ubuntu desktop. It restarts fine and gets to the place where I need to enter my password but after that it gets stock. I think it might not have done the swap file properly and the storage is now full so I am wondering if there is a way to reverse what I have done and try the installation again.

Thanks in advance!

Hello,

I build the tensorflow 1.3rc with the latest Jetpack 3.1.
cuDNN 6.0
CUDA 8.0
Python 2.7

thank you for the update!

Great thanks for your sharing. It works!

Tried simplified install:

  1. instal pip
  2. pip intstall whl
    The output is as follows:
nvidia@tegra-ubuntu:~/Desktop/tensorflow-tx2$ python
Python 2.7.12 (default, Nov 19 2016, 06:48:10) 
[GCC 5.4.0 20160609] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import tensorflow as tf
RuntimeError: module compiled against API version 0xb but this version of numpy is 0xa
RuntimeError: module compiled against API version 0xb but this version of numpy is 0xa

Shall try full install as it seems to me

To Andrey1984.
It seems like the version of numpy is different.
Please try to reinstall numpy.

sudo pip uninstall numpy
sudo pip install numpy

thank you for your response.
It was uninstalled with sudo pip uninstall nymphy
However it wont install at once

sudo pip install numpy-1.13.1
The directory '/home/nvidia/.cache/pip/http' or its parent directory is not owned by the current user and the cache has been disabled. Please check the permissions and owner of that directory. If executing pip with sudo, you may want sudo's -H flag.
The directory '/home/nvidia/.cache/pip' or its parent directory is not owned by the current user and caching wheels has been disabled. check the permissions and owner of that directory. If executing pip with sudo, you may want sudo's -H flag.
Collecting numpy-1.13.1
  Could not find a version that satisfies the requirement numpy-1.13.1 (from versions: )
No matching distribution found for numpy-1.13.1
nvidia@tegra-ubuntu:~$

Shall I determine a whl from there where I can not see aarch64:
https://pypi.python.org/pypi/numpy
?