Tensorflow 1.5.0
github.comThe big feature is CUDA 9 and cuDNN 7 support, which promises double-speed training on Volta GPUs/FP16. (it should be noted that TF 1.5 does not support CUDA 9.1 yet, which I found out the hard way)
I updated my Keras container with the TF 1.5 RC, CUDA 9, and cuDNN 7 (https://github.com/minimaxir/keras-cntk-docker), but did not notice a significant speed increase on a K80 GPU (I'm unsure if Keras makes use of FP16 yet either).
Speed in FP16 under k80 is almost as FP32, the architecture doesn't work well with FP16. As you said, you need Volta to notice improvements with FP16.
The other two main features are: Eager execution and TensorFlow Lite
You need Pascal hardware or later for FP16. Amazon's G3 Maxwell instances, which are newer than K80s, don't support it, either.
P3 instances have Volta (V100) GPUs.
k80 is pretty old now, 1080ti have x5 perf of k80 for one tenth of the price
When I say K80, I mean running a GPU in the cloud, not running it locally (and as an aside, due to the crypto boom, buying a physical GPU for cheap is more difficult)
A nice compromise is the GTX1080 that you can rent at Hetzner for 100e/month [0]. I am not sure how it compares to other GPUs but for experimenting it works quite well.
It's actually 118e/month (the price on the webiste are excluding VAT).
I'm using it, and I can only recommend - 0 problems, pretty decent box for a low price. Great solution, and waiting for the crypto bubble to burst and GPU prices to fall !
ah, 100e. it's still really good
that sounds really nice, thanks for the link
oh yeah. what i meant was that as k80 s are old, they don't have the fancy features like being fast with fp16, which came later. k80 come from slightly before the deep learning explosion afaik
Eager execution is appealing for folks new to learning TensorFlow. The deferred execution style is powerful, but if you just want to tinker in a REPL it's nice to have imperative programming. https://github.com/tensorflow/tensorflow/tree/r1.5/tensorflo...
What if I have data-dependent computational graphs? For example, recursive NNs.
I'm a member of the team that works on eager execution.
When eager execution is enabled, you no longer need to worry about graphs: operations are executed immediately. The upshot is that eager execution lets you implement dynamic models, like recursive NNs, using Python control flow. We've published some example implementations of such models on Github:
https://github.com/tensorflow/tensorflow/tree/master/tensorf...
I'd be happy to answer other questions about eager execution, and feedback is welcome.
EDIT: Just because you don't have to worry about graphs doesn't mean that graph construction and eager execution aren't related; take a look at our research blog post for more information if you're curious about the ways in which they relate to each other (https://research.googleblog.com/2017/10/eager-execution-impe...).
Looking forward to trying out eager!
I've been dreading version updates ever since they dropped Mac binary support. There are always obscure things to patch I have to find out by myself, the build easily wastes a whole day.
I think I'm either going to change my workflow and use another OS or switch fully to PyTorch.
I have recently compiled the -rc1 on MacOS 10.12, and the process has basically remained unchanged from 1.4. Here's a gist with my notes: https://gist.github.com/MatthiasWinkelmann/41dd42b07f8bd794a...
(I still used CUDA 8, but 9 should also work. You just need to find the version of the command line tools it works with)
And zero mention of AMD support.
Until there's something like CUDA for AMD isn't this going to be difficult? I don't know the area well, just curious.
(How good is openCL when it comes to this sort of stuff? Could they support it without crazy effort?)
AMD is porting TensorFlow, but these release notes give the impression that upstream is not helping in any way. https://github.com/ROCmSoftwarePlatform/hiptensorflow/blob/h...
For anyone who is having trouble with the installation, here's a tutorial to install TensorFlow 1.5.0 official pre-built pip package for both CPU and GPU version on Windows and ubuntu also there is tutorial to build tensorflow from source for cuda 9.1. http://www.python36.com
So is it an easy process to convert a tensorflow program from 1.4 to 1.5? When I tried converting something from 0.9 to 1.0 I couldn't figure it out.
For most cases it should just be a drop in replacement. IIRC they promise not to break the API between point releases (except tf.contrib.* which may change or disappear entirely...)
That promise is only for the Python API, not C++/Java.
0.9 wasn't production ready and they didn't guarantee backward compatibility until 1.0. So nothing to change from 1.4 to 1.5, maybe you will have some warnings about features that will change in the future, but it will work.
> Starting from 1.6 release, our prebuilt binaries will use AVX instructions. This may break TF on older CPUs.
This is primarily pre-2011 CPUs, though? Looks like everything since Sandy Bridge and Bulldozer will be okay.
I guess we never know what's running on our cloud instances.
IIRC since the beginning tensorflow has required at least sm 3.0 support (Kepler or newer). I imagine the combination of a pre-AVX cpu and Kepler or newer gpu is uncommon.