Keras Version Outdated

[ntungare]

The Keras version hosted is currently outdated. Is there going to be a new release soon?

Reference:
https://github.com/fchollet/keras#keras-deep-learning-library-for-tensorflow-and-theano

[ericdill]

@Ninad998 it is straightforward to do the update. Would you like to contribute the necessary changes to the recipe to build the new version of keras? I'm happy to walk you through the process.

[ntungare]

@ericdill I am a little busy now a days so I'm not too sure if I will have the time, I was hoping to use the latest Keras and found that the 1.1.0 is under a pull request but the build is failing on that.
#1

[ericdill]

Alright no worries. At any point in the future if you'd like me to walk you through the process, just ping me.

Thanks for reminding me about #1. I'll see if I can't get that building

[ghego]

Hey guys, I opened #4 to update to the most recent version in keras but it seems to have trouble running the ci suite. @ericdill can you have a look at what's going on?

[jakirkham]

I think the issue was we needed a newer tensorflow. Maybe wait until that is fully released and then give it another go.

[ghego]

Tensorflow 0.11 has just been merged in conda-forge/tensorflow-feedstock#22
However I'm afraid the issue has to do with this: conda-forge/tensorflow-feedstock#19, which seems to be still open. The problem is related to the build system used to run the tests, not to the update itself. Can anyone of the maintainers of the CI system fix it?

[ericdill]

@ghego rebase your PR and push it again to trigger the CIs again now that new tensorflow is available on anaconda.org/conda-forge

[jakirkham]

The problem is related to the build system used to run the tests, not to the update itself. Can anyone of the maintainers of the CI system fix it?

Not sure what you are proposing. Bumping the GLIBC version in the image? That's not exactly trivial and would be a larger org discussion. Not saying it is impossible, but it is not the easiest path forward and is likely to run into obstacles.

We have gotten keras to work with this binary repackaging of tensorflow before even with the older GLIBC use at conda-forge. I would think it should be possible again. One will just need a newer GLIBC to use either of these. Have restarted the CIs on that PR to see.

If that's not working, it might be worth trying to build tensorflow from source again. Something we want to do anyways. A fair bit of work has already gone into that.

[ntungare]

Hi,
There is still an issue with the Keras version on Conda for Windows 64-bit Python 2.

Using the syntax from Conda Cloud website.

https://anaconda.org/conda-forge/keras

[croth1]

I had troubles building keras on windows. Unfortunately, I don't have access to a windows machine to solve the problems. I am dependent on a windows savvy person to help me get the build working.

See the discussion in #6.

[jakirkham]

I think the problem should be resolvable if you are willing to contribute @Ninad998. 😉

[ntungare]

Actually I think the problem should be solved on Keras's end. Since the default requirement is Tensorflow and Tensorflow does not have a Python2 PyPi install for Windows.

I will raise an issue on Keras Github.

[jakirkham]

Ok, let us know how it goes.

Though still open to a PR that does the same thing on Windows that we do on Unix, should it be needed.

[MyVanitar]

You can use conda install -c hesi_m keras to install the latest version of Keras (2.1.6) and Tensorflow(1.8) for you. Unfortunately some Anaconda packages are outdated and people have to install them by the Pip which causes package conflicts with the conda and pip.

[jakirkham]

We are currently on 2.1.5 and there is a PR open for 2.1.6. The community is welcome to push this forward and take a more active roll in this and other feedstocks if they are of interest.