Merge this branch with Keras 2.0

[rravu3]

Hi,

Wanted to ask what are the plans to pull in Keras 2.0 as there have been interface level changes from 1.2.2.

Rahul

[kevinthesun]

Yes. We will first finalize keras/mxnet for 1.2.2 and then switch to 2.0.

[yliu7366]

Waiting for MXNet + Keras 2. I have networks developed based on Keras 2 and I am desperate for MXNet auto multi-GPU support....

[TimZaman]

@kevinthesun timeline? Why not try to merge this into Keras?

[edmondja]

:'( same as @yliu7366 I hope at least somebody has begun to work on this

[davideboschetto]

It would be great to have an update: are things still planned for a Keras2.0 upgrade?

[kevinthesun]

@sandeep-krishnamurthy

[sandeep-krishnamurthy]

Hi Yes. This is planned and work is in progress. I have implemented all basic tensor operators.. https://github.com/simpledl/keras/tree/mxnet_backend_keras2
Note above repo is just a placeholder I started working on. All work we do on Keras 2 will go to a PR on Keras 2.

[davideboschetto]

This is great @sandeep-krishnamurthy !
Keep up the good work, having the backend in keras2 would be amazing, thanks!
Do you by any chance have a ballpark estimation for when the PR will come?

[sandeep-krishnamurthy]

Hello,
Keras2+MXNet is almost complete. https://github.com/deep-learning-tools/keras/tree/keras2_mxnet_backend . Please try and provide your feedback.

I have created a PR - keras-team#9291
However, there are few design incompatibility with MXNet and Keras in the way Model compile and Optimizers are handled. Hence, few changes are requested by Keras community. These changes require significant modifications in MXNet Module and KVStore components. Hence, we talked to keras community to release the fork for the benefit of Keras users as the interface remains the same.

[edmondja]

@sandeep-krishnamurthy How to install the Keras2-compatible version ?

[sandeep-krishnamurthy]

@edmondja - Thanks for you interest.
Please open issue at - https://github.com/deep-learning-tools/keras if you face any problems.

$ mkdir keras
$ git clone https://github.com/deep-learning-tools/keras ./keras/
$ cd keras/

# MXNet backend code is in keras2_mxnet_backend branch. Check it out.
$ git fetch
$ git checkout keras2_mxnet_backend
$ git pull

# Install Keras
$ sudo python setup.py install

[edmondja]

Hello, I used all of your instructions but I have this error :

File "/Users/edmondjacoupeau/anaconda/lib/python3.6/site-packages/keras/backend/init.py", line 72, in
assert _backend in {'theano', 'tensorflow', 'cntk'}

AssertionError

[sandeep-krishnamurthy]

Hi @edmondja It looks like you have not checked out keras2_mxnet_backend branch. Can you please confirm?
Because, in init.py you should see assert_backend in {'theano', 'tensorflow', 'cntk', 'mxnet'} like this
https://github.com/deep-learning-tools/keras/blob/keras2_mxnet_backend/keras/backend/__init__.py#L72

[sandeep-krishnamurthy]

@edmondja - Did you had pre-installed keras (without mxnet backend)? If yes, please uninstall it before running above provided command.

$ pip uninstall keras

[edmondja]

Thanks it worked