TensorFlow 0.7.0

[Martin Wicke]

We are happy to announce that we are releasing TensorFlow 0.7.0 today. 

The release includes these major features:

• Allow using any installed Cuda >= 7.0 and cuDNN >= R2, and add support for cuDNN R4
• Added a contrib/ directory for unsupported or experimental features, including higher level layers module
• Added an easy way to add and dynamically load user-defined ops
• Built out a good suite of tests for the open-source version, things should break less!
• Added MetaGraphDef which makes it easier to save graphs with metadata
• Added assignments for "Deep Learning with TensorFlow" udacity course

It also contains over two months worth of new features, improvements and bug fixes, with many contributions by the community.

The full release notes can be found here, and installation instructions are here.

Enjoy!

The TensorFlow team

[Mohammed AlQuraishi]

Note that download link on master README.md still points to 0.6.

[Martin Wicke]

Mohammed, 

master is going to be updated by PR #1131, merging 0.7 back into master. As soon as the tests pass, I'll merge it.

Martin

--
You received this message because you are subscribed to the Google Groups "Discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to discuss+u...@tensorflow.org.
To post to this group, send email to dis...@tensorflow.org.
To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/discuss/39822767-6053-4196-a88c-3633523f56fb%40tensorflow.org.

[benba...@gmail.com]

Is the pip package still only built for cudnn v2 (i.e., is compiling from source necessary to use v4)? I get one of these two errors when trying to use it with v4:

F tensorflow/stream_executor/cuda/cuda_dnn.cc:693] failed to enqueue convolution on stream: CUDNN_STATUS_BAD_PARAM

F tensorflow/stream_executor/cuda/cuda_dnn.cc:828] failed to enqueue convolution on stream: CUDNN_STATUS_BAD_PARAM

[John Novak]

Are there any thoughts or plans to developing an Nvidia TK1/TX1
friendly distribution?

This is, of course, on the opposite end of the spectrum of large scale
deployment systems, but there are a lot of academic labs that can
afford a TX1 more easily than they can afford a dedicated machine.

> --
> You received this message because you are subscribed to the Google Groups
> "Discuss" group.
> To unsubscribe from this group and stop receiving emails from it, send an
> email to discuss+u...@tensorflow.org.
> To post to this group, send email to dis...@tensorflow.org.
> To view this discussion on the web visit

> https://groups.google.com/a/tensorflow.org/d/msgid/discuss/CADtzJKN0RC_PbPNyek2_Sx-HtH7Lxbcky_6nQaneNt5HLXQWUg%40mail.gmail.com.



--
John S. Novak, III
John....@gmail.com

[Geoffrey Irving]

This is unrelated to the 0.7.0 release, and should probably be filed as a Github issue instead.

To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/discuss/CAM3Ds1FEq1KvoW3Z5N%3DsHLyZNoW-THh2CoLRXF8UtsMPpu3ydA%40mail.gmail.com.

[zio...@gmail.com]

I'm getting the same error - F tensorflow/stream_executor/cuda/cuda_dnn.cc:693] failed to enqueue convolution on stream: CUDNN_STATUS_BAD_PARAM  - did you find any solution? I'm runing 0.70, with CUDA 7.0 (doagraded from 7.5 since it did not work)

The TensorFlow team

[zio...@gmail.com]

downgardaed to Cudnn v3 and it works for me!

On Wednesday, February 17, 2016 at 1:49:27 PM UTC-5, benba...@gmail.com wrote:

On Wednesday, February 17, 2016 at 1:49:27 PM UTC-5, benba...@gmail.com wrote:

On Wednesday, February 17, 2016 at 1:49:27 PM UTC-5, benba...@gmail.com wrote:

[gga...@gmail.com]

TK1 support would be highly appreciated. I'm a college student right now and there's no way I can afford the TX1 board. I just ordered the TK1, but apparently the tensorflow build for that one has stopped working. I'm a little disappointed. Any word on some progress on this front?

[Shawn]

What does 

• Added an easy way to add and dynamically load user-defined ops

mean?

I have searched for a little bit in the document, stackoverflow and github issues. Could you point to some examples? Thanks!

[Martin Wicke]

+Manjunath Kudlur wrote this. The tutorial on adding an op has been updated with the new information. It's a lot easier.

--

You received this message because you are subscribed to the Google Groups "Discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to discuss+u...@tensorflow.org.
To post to this group, send email to dis...@tensorflow.org.

To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/discuss/51f672f5-d593-4920-9b1f-eb2f6e73baae%40tensorflow.org.

[Li Shuai]

Could a link be provided? I assumed it is https://www.tensorflow.org/versions/r0.7/how_tos/adding_an_op/index.html#adding-a-new-op, but did not find it there. Many thanks.

[Martin Wicke]

That is indeed the place where you find the information. In particular,  https://www.tensorflow.org/versions/r0.7/how_tos/adding_an_op/index.html#using-the-op-in-python shows how to load it dynamically once you've built your own op.

[Li Shuai]

Oh. I though it would be something that would add an op to some op that is built using the python API of tensorflow, which the case is not, and is a better way to load shared library. I think in the source code, there is something that registers python op. Seems to be `fact` something. I have not figured it out yet, and previously thought it was a tutorial on that. I have not got the need to write a c++ op yet. Thanks anyway :).

[Grigory A]

Docker install: 

is graph freeze util is out of CPU Binary image, however it is in source?