[May 11] LibreELEC(KODI) - 20190324 - Pine A64 (+) |[May 11] LibreELEC(KODI) - 20190324 - PINE A64-LTS / SOPINE | [May 09] NEMS Linux 1.5 - Build1 - Pine A64 (+) | [May 09] NEMS Linux 1.5 - Build 1 - Rock64 | [May 09] NEMS Linux 1.5 - Build 1 - PINE A64-LTS / SOPINE | [May 09] NEMS Linux 1.5 - Build 1 - RockPro |[April 23] Q4OS ver 2.7-r5 - 1080P Pinebook / Pinebook

Project Inspiration | Get Started | IRC Logs | Forum Rules/Policy


Tensorflow
#1
Wink 
Hello,

I recently got my pine64 and I'm quite happy with it.
I installed the Xenial Ubuntu and I intend to use it for light Machine Learning projects so I can keep it running, learning and predicting and I just fetch the information when I need them...

However, I'm having trouble installing Tensorflow ( https://www.tensorflow.org/versions/r0.9...stallation ) with the following error:
Code:
$ export TF_BINARY_URL=https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-0.9.0-cp27-none-linux_x86_64.whl
$ pip install --upgrade $TF_BINARY_URL
tensorflow-0.9.0-cp27-none-linux_x86_64.whl is not a supported wheel on this platform.


Has anyone successfully installed it? Did you build from source or got the pip install to work?
Reply
#2
Not sure if this is relevant for Python, but ...-linux_x86_64-... does not seem like the correct architecture for the binaries on the A64. Should be arm64 or aarch64 or sth.... if you don't find pre-built binaries for arm64, you probably have to start at https://github.com/tensorflow/tensorflow and built them from the source... it seems to have a Raspberry Pi target (32bit ofc) but you could probably copy that and switch the output architecture to ARM64...
Come have a chat in the Pine A64 IRC channel >>
Reply
#3
search for recent work with nvidia jetson tx1. might help you.
Reply
#4
Thx for the replies,

I'm currently compiling Tensorflow from source, I roughly followed this guide: https://github.com/samjabrahams/tensorfl...r/GUIDE.md but had to make some changes to build bazel (don't think I've done it the propper way, though...). I searched for the TX1, they faced the same problem with bazel and I'll look further into that to see if there's a better way (I'll post my results here later, if I'm successful).

Tensorflow is really slow building, some memory heavy operations using all my Pine64 memory and 2GB swap file.
14 hours into compiling and the current status is: [816 / 1,151]
Reply
#5
Are you building with all cores? Otherwise that doesn't sound too bad for progress...
Come have a chat in the Pine A64 IRC channel >>
Reply
#6
I'm not sure, I ran the following command:

bazel build -c opt --jobs 4 --verbose_failures //tensorflow/tools/pip_package:build_pip_package

Jobs 4 is a workaround to deal with out of memory, not sure, however, what the default is for the local_resources flag.


In my rMBP (I know, unfair comparison), it took me less than an hour to compile... (no job limit, though)
Reply
#7
The only bigger thing I build on the Pine are Linux kernels, those take about 17 minutes with 4 cores and a heatsink... if you dont have a heatsink the cores will get throttled after reaching 82C....
Come have a chat in the Pine A64 IRC channel >>
Reply
#8
(07-25-2016, 09:19 AM)xalius Wrote: The only bigger thing I build on the Pine are Linux kernels, those take about 17 minutes with 4 cores and a heatsink... if you dont have a heatsink the cores will get throttled after reaching 82C....


And you tell me that now?!

How do I check the temperature?
It doesn't seem hot (luckily I have the acrylic case so it is well ventilated)
Reply
#9
run /usr/local/sbin/pine64_health.sh
Come have a chat in the Pine A64 IRC channel >>
Reply
#10
Finally!
I successfully build it!

I'll write a small guide later on for those interested.
I'll also upload the final whl for the lazy ones... xD
Reply


Forum Jump:


Users browsing this thread: 1 Guest(s)