4 ms·

TensorFlow 1.7.0 released

165 points by kyloon 9y ago

m_ke 9y ago
For people who had time to play with TF Eager, how does it stack up vs pytorch?
Svenstaro 9y ago
Currently compiling this for Arch. Should be in the repos soon. :)
- sulam 9y ago
  It’s funny I was just downloading TF today and saw this. Compiled it for Ubuntu (took 40 minutes, too). Glad to see it since I think that means I can use CUDA 9.1.
- Yuioup 9y ago
  Compiling will go faster if you use a video card :-P
  - Svenstaro 9y ago
    No, it's actually much slower to compile with CUDA support because it has to compile all the CUDA kernels.
hcheemskerk 9y ago
Is there some sort of compatibility chart for tensorflow versions relating to CUDA/cuDNN? "Use the latest install guide" doth butter no parsnips if you're working with an older version of tensorflow.
- snoshy 9y ago
  https://www.tensorflow.org/install/install_sources#tested_source_configurations https://www.tensorflow.org/install/install_sources#tested_so...
  - hcheemskerk 9y ago
    Awesome, thanks!
oerpli 9y ago
Does someone know what's the current state of TF on Windows? I've used it for some image classification homework 1.5y ago and back then I had some issues but it kind of worked. Does someone have more recent experiences in some hobby project or even in a production env? Or are there other NN frameworks when using Windows?
- JPKab 9y ago
  It works fine on Windows now.
- aseipp 9y ago
  It's pretty easy to get running on Windows since 1.0 or whatnot, in my experience. I just ran `pip3 install --upgrade tensorflow-gpu` and do that on every release and it works just dandy. You get a binary wheel of course, so the install is quick and painless. I did have to install a new cudnn 7 DLL recently after moving to 1.5 (or 1.6) with CUDA 9. That's about it. I don't use all the fancier tools though (e.g. loading from GCS/S3 buckets etc, just local data for me). But everything seems to works just fine these days, it seems.
- markatkinson 9y ago
  Yea I managed to get it working pretty quickly and I was a complete newbie. Keras made it MUCH easier the second time around.
seesawtron 9y ago
Still with cuda9.0 and not cuda9.1
kyloon 9y ago
Official announcement from Google: https://developers.googleblog.com/2018/03/tensorrt-integration-with-tensorflow.html https://developers.googleblog.com/2018/03/tensorrt-integrati...
chanchar 9y ago
Anyone have a nice tl;dr?

165 points by kyloon 9y ago

m_ke 9y ago

For people who had time to play with TF Eager, how does it stack up vs pytorch?

Svenstaro 9y ago

Currently compiling this for Arch. Should be in the repos soon. :)

sulam 9y ago
It’s funny I was just downloading TF today and saw this. Compiled it for Ubuntu (took 40 minutes, too). Glad to see it since I think that means I can use CUDA 9.1.
Yuioup 9y ago
Compiling will go faster if you use a video card :-P
- Svenstaro 9y ago
  No, it's actually much slower to compile with CUDA support because it has to compile all the CUDA kernels.

hcheemskerk 9y ago

Is there some sort of compatibility chart for tensorflow versions relating to CUDA/cuDNN? "Use the latest install guide" doth butter no parsnips if you're working with an older version of tensorflow.

snoshy 9y ago
https://www.tensorflow.org/install/install_sources#tested_source_configurations https://www.tensorflow.org/install/install_sources#tested_so...
- hcheemskerk 9y ago
  Awesome, thanks!

oerpli 9y ago

Does someone know what's the current state of TF on Windows? I've used it for some image classification homework 1.5y ago and back then I had some issues but it kind of worked. Does someone have more recent experiences in some hobby project or even in a production env? Or are there other NN frameworks when using Windows?

JPKab 9y ago
It works fine on Windows now.
aseipp 9y ago
It's pretty easy to get running on Windows since 1.0 or whatnot, in my experience. I just ran `pip3 install --upgrade tensorflow-gpu` and do that on every release and it works just dandy. You get a binary wheel of course, so the install is quick and painless. I did have to install a new cudnn 7 DLL recently after moving to 1.5 (or 1.6) with CUDA 9. That's about it. I don't use all the fancier tools though (e.g. loading from GCS/S3 buckets etc, just local data for me). But everything seems to works just fine these days, it seems.
markatkinson 9y ago
Yea I managed to get it working pretty quickly and I was a complete newbie. Keras made it MUCH easier the second time around.

seesawtron 9y ago

Still with cuda9.0 and not cuda9.1

kyloon 9y ago

Official announcement from Google: https://developers.googleblog.com/2018/03/tensorrt-integration-with-tensorflow.html https://developers.googleblog.com/2018/03/tensorrt-integrati...

chanchar 9y ago

Anyone have a nice tl;dr?