Announcing TensorFlow.jl, an interface to Google's TensorFlow machine learning library

[Jonathan Malmaud]

Hello,

I'm pleased to announce the release of TensorFlow.jl, enabling modern GPU-accelerated deep learning for Julia. Simply run Pkg.add("TensorFlow") to install and then read through the documentation at https://malmaud.github.io/tfdocs/index.html to get started. Please file any issues you encounter at https://github.com/malmaud/TensorFlow.jl. 

TensorFlow.jl offers a convenient Julian interface to Google's TensorFlow library. It includes functionality for building up a computation graph that encodes a deep-learning model and automatically minimizing an arbitrary loss function with respect to the model parameters. Support is included for convolutional networks, recurrent networks with LSTMs, the Adam optimization algorithm, loading images, and checkpointing model parameters to disk during training

I'm hopeful that this package will ensure Julia remain a first-class citizen in world of modern machine learning and look forward to the community's help in getting it to match or exceed the capabilities of the official Python TensorFlow API. 

-Jon

[Jeffrey Sarnoff]

 This is very nice, Jon.

    (interested parties)

        regards,

         Jeffrey

[Kyunghun Kim]

Wonderful jobs, Jonathan! 

I'd better try this version rather than use TensorFlow in python. 

Does it based on PyCall package? 

-Kyunghun

2016년 9월 1일 목요일 오전 7시 31분 58초 UTC+9, Jonathan Malmaud 님의 말:

[Páll Haraldsson]

On Thursday, September 1, 2016 at 7:45:05 AM UTC, Kyunghun Kim wrote:

Wonderful jobs, Jonathan! 

I'd better try this version rather than use TensorFlow in python. 

Does it based on PyCall package?

Yes, you can always see that by looking in the REQUIRE file (except if used indirectly.. then you would have too look at packages recursively..).


I guess you can also see at runtime, by what library (meaning libpython, not general Julia packages, maybe strace would then help..) is used..

like [I know there's a command for it and this seems to be it]:

pldd <pid> |grep libpython


[see also lsof command.]

--
Palli.

[Jonathan Malmaud]

Thanks Kyunghun! It mostly uses the TensorFlow C library. It does rely on PyCall for now for the autodifferentiation functionality, which is not yet part of the C API. Google has said that the C API will soon expose AD functionality, at which point this package won't depend on Python at all.

[Páll Haraldsson]

On Wednesday, August 31, 2016 at 10:31:58 PM UTC, Jonathan Malmaud wrote:

Hello,

I'm pleased to announce the release of TensorFlow.jl, enabling modern GPU-accelerated deep learning for Julia. Simply run Pkg.add("TensorFlow") to install and then read through the documentation at https://malmaud.github.io/tfdocs/index.html to get started. Please file any issues you encounter at https://github.com/malmaud/TensorFlow.jl. 

About: "To enable support for GPU usage (Linux only)"

[I'm not complaining.. more asking about the general issue] It seems official Tensorflow supports (for *GPU*) OS X (but not Windows).

https://www.tensorflow.org/versions/r0.10/get_started/os_setup.html

Do you not support the same because you simply say do not have OS X or is this about some OS support missing from other packages?

>will ensure Julia remain a first-class citizen in world of modern machine learning

That would be great! You mean it's already ("remain") with other ANN packages (or yours, not your first version)? I guess people/you want Tensorflow (at least known in the Python world), I know nothing about it or the (Julia-native or otherwise..) competition.

--
Palli.

[Vishnu Raj]

Thanks Jon, I was planning to start learning Tensorflow - now I can do it in julia :)